A voice control-based intelligent wake-up system and method for household appliances

Abstract

The application discloses a kind of based on voice control's home appliance intelligent wake-up system and method, it is related to voice control technical field.This kind of based on voice control's home appliance intelligent wake-up system and method, comprising: S1, acquisition voice wake-up process observation data and obtain home appliance control execution data, carry out timing uniform, quality order and standardization expression;S2, based on link switching observation data carries out leading connection constraint analysis, carries out leading voice reservation length control and continuous connection speech segment generation;S3, to the boundary excursion of continuous connection speech segment, spectrum connection relationship and silence fracture state carry out recovery correction analysis;S4, carry out control semantic confirmation, control instruction issue and closed loop check-out.Solve the problem that existing method adopts hierarchical monitoring and time-sharing wake-up architecture, and it is easy to cause continuous voice first segment loss, response delay fluctuation and control execution incomplete due to link connection timing inconsistency in hibernation switching process.

Description

Technical Field

[0001] This invention relates to the field of voice control technology, specifically to a voice-controlled smart wake-up system and method for home appliances. Background Technology

[0002] With the continuous development of artificial intelligence, voice recognition, and smart home technologies, voice-based appliance control has gradually become one of the important human-computer interaction forms of smart terminals. In existing home appliances, voice control functions are widely used in air conditioners, televisions, lighting equipment, air purifiers, smart speakers, and various kitchen appliances. Users can start devices, switch modes, adjust parameters, and control linkages through voice wake-up, voice command input, and semantic parsing. To meet the needs of home appliances for long standby times, low power consumption, and immediate response, existing voice wake-up systems typically adopt an architecture combining terminal-side persistent monitoring with subsequent recognition and processing. Voice monitoring is maintained in standby mode, and subsequent recognition and control processes only begin after the target voice is detected. Simultaneously, existing technologies generally combine voice acquisition, noise suppression, feature extraction, keyword detection, voice recognition, semantic understanding, and control mapping to parse user voice content and convert the parsed results into control commands executable by the home appliances. Overall, building home appliance voice interaction capabilities around low-power standby, voice wake-up, command recognition, and control execution has become an important technological development direction in the current smart home and smart appliance fields.

[0003] For example, invention patent CN111179931B discloses a method, device, and household appliance for voice interaction. The method includes: echo cancellation of wake-up speech and acquisition of interference noise energy; interference removal energy processing of the echo-cancelled wake-up speech based on the interference noise energy; obtaining the average peak energy of the interference removal wake-up speech; and energy normalization processing of the average peak energy. Furthermore, it extracts the energy characteristics and power spectrum characteristics of the environmental noise after echo cancellation, combines them with an energy prediction model to predict the interference noise energy of different speech frames, and updates the energy prediction model based on the power spectrum characteristics. The invention also discloses a device and household appliance for voice interaction, which suppresses interference noise in wake-up speech, improving the accuracy of energy feature calculation and voice interaction effect in distributed voice interaction scenarios.

[0004] For example, the invention patent with publication number CN108022590B discloses a focusing session method for a voice interface device. This method is applied to a first electronic device in a local group of connected electronic devices, and includes: receiving a first voice command including a request for a first operation; determining a first target device from the local group for the first operation; establishing a focusing session relative to the first target device and having the first operation executed by the first target device; receiving a second voice command including a request for a second operation; determining that the second operation can be executed by the first target device when the second voice command does not explicitly specify a second target device, and judging whether the second voice command satisfies a focusing session maintenance criterion; and, when the focusing session maintenance criterion is satisfied, having the second operation continue to be executed by the first target device. This achieves continuous orientation and execution of the target device for subsequent voice commands when the target device is not explicitly specified or the specification is ambiguous, improving the coherence and accuracy of voice control in a multi-device environment.

[0005] While existing home appliance voice interaction technologies can improve interaction effects through noise suppression, voice wake-up recognition, and continuous target device orientation, most related solutions focus on anti-interference processing of wake-up voice, improving command recognition accuracy, or target device matching in multi-device scenarios. They do not pay enough attention to the continuity of voice during the switching process from low-power monitoring link to main processing link. In particular, they lack collaborative processing mechanisms for preamble voice retention after wake-up trigger, link handover timing coordination, splicing boundary correction, and control execution closure confirmation. As a result, in hierarchical monitoring and time-sharing wake-up architecture, the information in the preceding part of continuous voice is easily lost during sleep switching, which affects the completeness of subsequent control statement recognition, response timing consistency, and control execution reliability.

[0006] Therefore, in order to address the above problems, there is an urgent need for a voice-controlled smart wake-up system and method for home appliances. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a voice-controlled smart wake-up system and method for home appliances. It solves the problems of loss of continuous voice segments, response delay fluctuations, and incomplete control execution caused by inconsistent link connection timing during the sleep switching process when existing methods adopt a hierarchical monitoring and time-sharing wake-up architecture.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a voice-controlled smart wake-up method for home appliances, comprising: S1, collecting observation data of the voice wake-up process and obtaining home appliance control execution data, and performing time-series unification, quality ordering, and standardized expression on the voice wake-up process observation data and home appliance control execution data; S2, performing preamble constraint analysis based on link switching observation data, and controlling the preamble speech retention length and generating continuous speech segments based on the preamble constraint analysis results; S3, performing restoration and correction analysis on the boundary offset, spectral connection relationship, and silence break state of continuous speech segments, and performing wake-up word end position correction, control statement start position correction, and continuous speech reconstruction; S4, performing execution closure analysis in conjunction with home appliance control execution data, and performing control semantic confirmation, home appliance control command issuance, and execution result closed-loop verification based on the execution closure analysis results.

[0011] Furthermore, the specific process of collecting observation data of the voice wake-up process and obtaining home appliance control execution data is as follows: In the home appliance voice wake-up terminal, the voice wake-up process observation data and home appliance control execution data are collected. The voice wake-up process observation data includes: continuously collecting indoor raw voice waveform data through analog-to-digital conversion sampling circuit and microphone array, and synchronously recording the voice sampling timestamp sequence; real-time reading of the replayable voice retention duration data, low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, voice recognition model loading completion timestamp, and continuous voice takeover start timestamp in the circular voice buffer; obtaining home appliance control execution data: obtaining home appliance identification, home appliance type data, and home appliance executable instruction set data through the local device registry and local control protocol table; reading the home appliance control instruction issuance timestamp, home appliance execution receipt timestamp, and home appliance execution receipt status data through the control bus log and execution receipt log.

[0012] Furthermore, the specific process for unifying the timing, quality ordering, and standardizing the voice wake-up process observation data and home appliance control execution data is as follows: Unify the timestamp alignment and remove duplicate records from the voice wake-up process observation data and home appliance control execution data; perform isolated spike suppression processing on the original speech waveform data through amplitude mutation detection at adjacent sampling points; perform missing record completion and protocol consistency verification processing on the home appliance control execution data; extract the time-spectrum sequence from the original speech waveform data through short-time Fourier transform processing, and extract the log-Mel spectrum sequence from the time-spectrum sequence through log-Mel spectrum extraction processing; standardize the log-Mel spectrum sequence, voice wake-up process observation data, and home appliance control execution data using the Z-score normalization algorithm; and normalize the low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, speech recognition model loading completion timestamp, continuous voice takeover start timestamp, home appliance control command issuance timestamp, and home appliance execution receipt timestamp using the maximum and minimum value normalization algorithm.

[0013] Furthermore, the specific process of performing preamble retention constraint analysis based on link switching observation data is as follows: Obtain the retrievable voice retention duration data, low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, speech recognition model loading completion timestamp, and continuous voice takeover start timestamp; calculate the sum of the first and ring voice buffer write durations, take the arctangent of the result to obtain the preamble retention compression term; calculate the difference between the speech recognition model loading completion timestamp and the main clock lock timestamp, calculate the difference between the main clock lock timestamp and the main processing power-on completion timestamp, calculate the product of the two differences, take the inverse hyperbolic sine and add one to obtain the link establishment suppression term; calculate the difference between the continuous voice takeover start timestamp and the low-power wake-up trigger timestamp, divide it by the sum of the first and ring voice buffer write durations, and take the negative exponent to obtain the retention hysteresis attenuation term; divide the preamble retention compression term by the link establishment suppression term, and multiply it by the retention hysteresis attenuation term to obtain the preamble retention constraint value.

[0014] Furthermore, the specific process for controlling the preamble retention length and generating continuous speech segments based on the preamble retention constraint analysis results is as follows: Real-time comparison of the preamble retention constraint value and the preamble retention constraint threshold: When the preamble retention constraint value is greater than or equal to the preamble retention constraint threshold, the length of the preamble already retained in the circular speech buffer is maintained. The preamble speech segment is read back from the circular speech buffer according to the preamble readback duration n before the low-power wake-up trigger timestamp. The read-back preamble speech segment is then combined with the newly sampled speech segment after the continuous speech takeover start timestamp to form a continuous speech segment. When the preamble retention constraint value is less than the preamble retention constraint threshold, the retention duration of the circular speech buffer is extended to maintain synchronous retention of the retention time between the low-power monitoring link and the main processing link. The speech data is processed within the interval; during the retention time interval, dual-path mirror writing is performed on the leading speech segment and the newly sampled speech segment, so that the leading speech segment and the newly sampled speech segment are retained in the corresponding retention interval of the circular speech buffer and the corresponding takeover interval of the main processing link, respectively. Then, the splicing starting point is re-trimmed based on the low-power wake-up trigger timestamp, the master clock lock timestamp, and the continuous speech takeover start timestamp. The position corresponding to the continuous speech takeover start timestamp is used as the truncation benchmark, and the truncation position offset correction is performed in combination with the low-power wake-up trigger timestamp and the master clock lock timestamp. The truncated speech segments are re-weighted and spliced ​​in the overlapping area with continuously changing weighting coefficients to fill the speech gaps. After the compensation is completed, a continuous speech segment is generated, and then it enters the continuous speech reconstruction and boundary correction processing.

[0015] Furthermore, the specific process for restoring and correcting the boundary offset, spectral coherence, and silence breakage states of consecutively connected speech segments is as follows: The wake-up tail rewind amount is calculated from the log-Mel spectrum sequence of the consecutively connected speech segments using a wake-up word template dynamic time warping algorithm; the segment-first spectral coherence degree is calculated from the log-Mel spectrum sequences on both sides of the connection position formed when the preceding speech segment and the newly sampled speech segment in the consecutively connected speech segment are processed using spectral cosine similarity; and the original speech waveform data and speech sampling timestamp sequence of the consecutively connected speech segments are analyzed using energy statistics of adjacent speech analysis intervals and joint detection of zero-crossing rates. The method involves measuring and identifying silent discontinuities where energy continuously decreases and the zero-crossing rate changes synchronously. The density of broken silence is then calculated based on the frequency and duration of these discontinuities per unit time. The integral value of the sum of one divided by one and the square of the integral variable within the interval from zero to the wake-up tail segment is calculated. Taking the negative exponent of the integral value yields the boundary offset suppression term. The error function value of the spectral connectivity at the beginning of the segment is calculated, and one is added to obtain the spectral connectivity enhancement term. The sum of one and the broken silence density is calculated, and the natural logarithm is taken and one is added to obtain the silent break suppression term. Finally, the boundary restoration reliability value is obtained by multiplying the boundary offset suppression term by the spectral connectivity enhancement term and dividing by the silent break suppression term.

[0016] Furthermore, the specific process for correcting the end position of the wake-up word, the start position of the control statement, and the continuous speech reconstruction is as follows: Real-time comparison of the boundary restoration confidence value and the boundary restoration confidence threshold: When the boundary restoration confidence value is greater than or equal to the boundary restoration confidence threshold, the path endpoint obtained by performing wake-up word template dynamic time warping matching on the log-Mel spectrum sequence of the continuous speech segment is used as the wake-up tail truncation position. The continuous speech segment is divided into wake-up word segments and control statement segments according to the wake-up tail truncation position; silence boundary re-determination is performed on the segmented control statement segments, and the initial retained sample points before the first phonation point are retained. The reconstructed complete control statement fragment is then used for password recognition. When the boundary restoration confidence value is less than the boundary restoration confidence threshold, restoration processing is performed on the continuous speech segments: the connection position formed during the double-end splicing process of the newly sampled speech segment is pushed back according to the wake-up tail segment back amount; the silence intervals in the adjacent speech analysis intervals on both sides of the connection position are deleted or merged according to the break silence density; the connection correction interval is elastically resampled in the spectral connection interval on both sides of the connection position according to the segment start spectral connection degree; after the processing is completed, the corrected wake-up word end position and control statement start position are determined, and then password recognition is performed.

[0017] Furthermore, the specific process of performing execution closure analysis based on home appliance control execution data is as follows: For complete control statement fragments, the posterior probability sequence of the command is obtained through a speech recognition model, and then the posterior set value of the command is obtained through Gaussian cumulative distribution function mapping; For complete control statement fragments, the device reference text is obtained by decoding through a speech recognition model and combining it with device word extraction rules; The device reference text and home appliance type data are then compared using an edit distance normalization algorithm to calculate the device word matching degree; For complete control statement fragments and home appliance executable instruction set data, the instruction slot missing degree is calculated using an intent slot filling algorithm; The execution receipt closure time difference is calculated by the difference between the home appliance control instruction issuance timestamp and the home appliance execution receipt timestamp; The product of the posterior set value of the command and the sum of the device word matching degree and one is calculated to obtain the command-device coordination term; The square root of the sum of the square of the instruction slot missing degree and the square of the execution receipt closure time difference is calculated, and one is added to obtain the execution damping term; The command-device coordination term is divided by the execution damping term, one is added, and the natural logarithm is taken to obtain the command execution confidence value.

[0018] Furthermore, the specific process of confirming control semantics, issuing home appliance control commands, and verifying execution results based on the results of closed-loop analysis is as follows: Real-time comparison of the password execution credibility value and the password execution credibility threshold; when the password execution credibility value is less than the password execution credibility threshold, restricted word list error correction decoding and slot completion re-parsing are performed on the complete control statement fragment; for cases where the device reference text cannot uniquely correspond to a single home appliance type, nearest neighbor matching is re-performed in the home appliance type data, and the decoding space is constrained by the re-matched device set; for cases where the slot filling result corresponding to the missing command slot does not cover the target action slot and target parameter slot in the current executable command set of the home appliance, the missing actions and parameters are filled in according to the executable command set data of the home appliance. After re-decoding, the password execution confidence value is recalculated. If the password execution confidence value is still less than the password execution confidence threshold, only a local voice confirmation prompt is output, and no control is issued. When the password execution confidence value is greater than or equal to the password execution confidence threshold, candidate control items consistent with the current device are selected from the executable instruction set data of the home appliance based on the home appliance type data, and the complete control statement fragment is subjected to secondary decoding using a restricted word list. After secondary decoding, a unique home appliance control instruction is generated and issued to the target device corresponding to the home appliance identifier through the control bus. After issuance, the home appliance execution receipt status data and the home appliance execution receipt timestamp are continuously read. The control instruction and execution receipt are matched one-to-one to confirm whether the current round of control is closed. When the verification is consistent, the final execution result is output.

[0019] The second aspect of this invention provides a voice-controlled smart home appliance wake-up system, comprising: a data acquisition and preprocessing module, used to acquire observation data of the voice wake-up process and obtain home appliance control execution data, and to perform time-series unification, quality ordering, and standardized expression on the voice wake-up process observation data and home appliance control execution data; a preamble constraint discrimination module, used to perform preamble constraint analysis based on link switching observation data, and to control the preamble speech retention length and generate continuous speech segments based on the preamble constraint analysis results; a continuous speech reconstruction and boundary correction module, used to perform restoration and correction analysis on the boundary offset, spectral connection relationship, and silence break state of continuous speech segments, and to perform wake-up word end position correction, control statement start position correction, and continuous speech reconstruction; and a command recognition and execution closure module, used to perform execution closure analysis in conjunction with home appliance control execution data, and to perform control semantic confirmation, home appliance control command issuance, and execution result closed-loop verification based on the execution closure analysis results.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention, by constraining and judging the preamble speech retention length and link connection relationship, enables the preamble speech segment after wake-up triggering to be effectively retained and utilized in subsequent processing, avoiding the preamble of the wake-up word being covered or omitted during link establishment, thereby achieving the effect of stable retention of wake-up preamble speech, effectively solving the problem of preamble speech being difficult to completely preserve under the hierarchical monitoring architecture in the prior art.

[0023] (2) This invention restores and corrects the boundary offset, spectrum connection relationship and silence break state of continuous speech segments, making the boundary position between the wake-up word segment and the control statement segment clearer, avoiding the inaccurate segmentation of control statements caused by splicing misalignment and boundary drift, thereby achieving the effect of accurate reorganization of speech boundary, effectively solving the problem that the boundary between the wake-up word and the control statement is difficult to accurately restore after speech splicing in the prior art.

[0024] (3) This invention, by performing restricted word list decoding, device correspondence confirmation and slot completion re-parsing on control statement fragments, enables the speech recognition results to form a more stable mapping relationship with the control semantics of home appliances, avoiding semantic incompleteness or object ambiguity from directly entering the control execution stage, thereby achieving the effect of reliable confirmation of control semantics, and effectively solving the problem that the incomplete recognition of control statements in the prior art can easily lead to erroneous control.

[0025] (4) This invention introduces the execution receipt closed state into the voice control result determination process, so that the generation of control instructions, the issuance of control instructions and the confirmation of execution results form a closed-loop verification relationship, avoiding the direct determination of control success based solely on the voice recognition result, thereby achieving the effect of verifiable control execution process, and effectively solving the problem of lack of execution result closed confirmation after voice control is completed in the prior art.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 This is a flowchart of a voice-controlled smart wake-up method for home appliances according to the present invention.

[0028] Figure 2 This is a structural diagram of a voice-controlled smart wake-up system for home appliances according to the present invention.

[0029] Figure 3 This is a density distribution and confidence elliptic plot of the leading acceptance constraint value of the present invention;

[0030] Figure 4 This is a bar chart comparing the preceding acceptance constraint values ​​of the present invention;

[0031] Figure 5This is a flowchart of the boundary restoration confidence value determination and speech segment processing of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figures 1-5 This invention provides a technical solution: a voice-controlled smart wake-up method for home appliances, comprising the following steps: S1, collecting observation data of the voice wake-up process and obtaining home appliance control execution data, and performing time-series unification, quality ordering, and standardized expression on the voice wake-up process observation data and home appliance control execution data; S2, performing preamble constraint analysis based on link switching observation data, and controlling the preamble speech retention length and generating continuous speech segments according to the preamble constraint analysis results; S3, performing restoration and correction analysis on the boundary offset, spectral connection relationship, and silence break state of continuous speech segments, and performing wake-up word end position correction, control statement start position correction, and continuous speech reconstruction; S4, performing execution closure analysis in conjunction with home appliance control execution data, and performing control semantic confirmation, home appliance control command issuance, and execution result closed-loop verification according to the execution closure analysis results.

[0034] Specifically, the process of collecting observation data of the voice wake-up process and obtaining home appliance control execution data is as follows: In the home appliance voice wake-up terminal, observation data of the voice wake-up process and home appliance control execution data are collected. The collection of observation data of the voice wake-up process includes: continuously collecting indoor raw voice waveform data through analog-to-digital conversion sampling circuit and microphone array, and synchronously recording the voice sampling timestamp sequence; reading the replayable voice retention duration data in the ring voice buffer in real time through the ring buffer control unit; reading the low-power wake-up trigger timestamp through the low-power monitoring control register, which is the time stamp formed when the low-power monitoring link outputs the wake-up judgment; reading the main processing power-on completion timestamp, the main clock lock timestamp, and the voice recognition model loading completion timestamp through the main processing unit power management register, clock lock register, and model loading status register, respectively. The main processing power-on completion timestamp is the time stamp formed when the main processing link enters the startable state, and the main clock lock timestamp is the time stamp formed when the main processing link enters the startable state. The fixed timestamp is a time stamp formed when the main processing link clock is established, and the timestamp for the completion of speech recognition model loading is a time stamp formed when the speech recognition computing resources are loaded. The continuous speech takeover start timestamp is read through the speech takeover status register, which is a time stamp formed when the main processing link starts receiving continuous speech data. Home appliance control execution data is obtained: home appliance identifier, home appliance type data, and home appliance executable instruction set data are obtained through the local device registry and local control protocol table. The timestamps of home appliance control command issuance, home appliance execution receipt timestamps, and home appliance execution receipt status data are read through the control bus log and execution receipt log. The speech sampling timestamp sequence, low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, speech recognition model loading completion timestamp, continuous speech takeover start timestamp, home appliance control command issuance timestamp, and home appliance execution receipt timestamp are recorded according to a unified time base.

[0035] This implementation plan achieves unified collection, unified calibration, and unified timing organization of observation data of the voice wake-up process and home appliance control execution data. This enables key state changes in the low-power monitoring link, main processing link, and control execution link to form a corresponding, traceable, and verifiable data foundation under the same time reference. It improves the overall consistency and comparability among the preamble voice retention state, link switching establishment state, continuous voice takeover state, and control execution feedback state. It also enhances the ability of subsequent preamble acceptance constraint discrimination, boundary restoration reliability judgment, and command execution closure verification to identify cross-stage timing deviations, state drifts, and control mismatches. This provides stable data support for the generation of continuous acceptance voice segments, complete reconstruction of control statements, and accurate issuance of home appliance control commands.

[0036] Specifically, the process of unifying the timing, quality ordering, and standardizing the expression of voice wake-up process observation data and home appliance control execution data is as follows: The voice wake-up process observation data and home appliance control execution data undergo unified timestamp alignment and duplicate record removal; asynchronous records from the voice acquisition link, link switching link, and control execution link form a corresponding timing sequence under the same time reference, reducing the impact of timing drift during cross-stage state comparison; isolated spike suppression is performed on the original voice waveform data through amplitude mutation detection at adjacent sampling points; non-voice pulse components introduced by instantaneous electromagnetic disturbances, sampling jitter, or sudden input abnormalities are suppressed to avoid local distortion of subsequent time-spectrum expansion and boundary determination caused by abnormal spikes; missing record completion and protocol consistency verification are performed on the home appliance control execution data; the semantic correspondence between control-side records in terms of device type, command expression, and execution receipt is maintained, reducing execution closure deviation caused by missing control protocol items or inconsistent state expressions; the time-spectrum sequence is extracted from the original voice waveform data through short-time Fourier transform processing, and then further processed through log-Mel spectrum extraction. Log-Mel spectrum sequences are used to ensure a unified spectral carrier for subsequent wake-up boundary localization and command reconstruction; the original speech waveform data is transformed from a time-domain representation to a spectral representation that takes into account both time distribution and frequency band response, enhancing the ability to represent changes in the onset of speech, splicing boundary disturbances, and syllable connection differences; the log-Mel spectrum sequences, voice wake-up process observation data, and home appliance control execution data are standardized using the Z-score standardization algorithm; the numerical span differences caused by different dimensions, sampling scales, and recording ranges are removed, improving the comparability of various input quantities and the stability of joint calculation in subsequent discrimination processes; the low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, speech recognition model loading completion timestamp, continuous voice takeover start timestamp, home appliance control command issuance timestamp, and home appliance execution receipt timestamp are normalized using the maximum and minimum value normalization algorithm to ensure that the relevant parameters in subsequent formula calculations are dimensionless; the time series quantities corresponding to each key time point participate in the calculation of constraint values ​​and confidence values ​​within a unified numerical range, reducing the amplification effect of absolute time scale differences on the formula output results.

[0037] This implementation plan achieves unified cleaning, transformation, and dimensional organization of observation data and home appliance control execution data during the voice wake-up process. This enables multi-source asynchronous data in the voice acquisition link, link switching link, and control execution link to participate in subsequent processing based on consistent, stable, and comparable data. It reduces the interference of abnormal spikes, missing records, inconsistent protocol expressions, and differences in absolute time scales on subsequent judgment results. It enhances the identifiability and computability of changes in the onset of voice production, splicing boundary disturbances, syllable connection differences, and control execution closed states. It provides stable, standardized, and jointly computable data support for calculating the preceding constraint value, boundary restoration confidence value, and command execution confidence value.

[0038] Specifically, the process of conducting pre-transfer constraint analysis based on link switching observation data is as follows: acquire the replayable voice retention duration data, low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, speech recognition model loading completion timestamp, and continuous voice takeover start timestamp. The replayable voice retention duration data is used to characterize the continuous retention range of replayable voice samples during the low-power monitoring phase. The low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, speech recognition model loading completion timestamp, and continuous voice takeover start timestamp constitute a transfer timing sequence according to the link startup order, which is used to characterize the segmented establishment process from wake-up trigger to voice takeover completion.

[0039] The sum of the writing duration of the first and the circular speech buffer is calculated, and the arctangent of the result is taken to obtain the preamble retention compression term. The arctangent function is used to perform compression mapping on the writing duration of the circular speech buffer, making the incremental change of the writing duration within a large value range more gradual, thus avoiding the preamble retention capability from dominating the formula output too strongly. The difference between the speech recognition model loading completion timest and the master clock locking timest, and the difference between the master clock locking timest and the main processing power-on completion timest are calculated. The product of these two differences is then multiplied, and the inverse hyperbolic sine is added to obtain the link establishment suppression term. The product of the time differences of two adjacent stages is then used to obtain the link establishment suppression term. The cascading lag during link establishment is characterized, and the product order is compressed using an inverse hyperbolic sine function. Adding one term keeps the denominator positive, allowing related timing quantities to participate in constraint calculations in dimensionless form. The difference between the continuous voice takeover start timestamp and the low-power wake-up trigger timestamp is calculated, divided by the sum of the calculation and the circular voice buffer write duration, and then a negative exponent is taken to obtain the lag attenuation term. The difference is used to form a normalized ratio with the circular voice buffer write duration, and a negative exponent mapping is used to highlight the attenuation effect of lag on preamble retention capability, keeping this term within a bounded interval. The preamble retention compression term is divided by the link establishment suppression term. Multiplying this by the hysteresis attenuation term yields the preamble acceptance constraint value, used to comprehensively characterize the preamble retention capability, link establishment time, and hysteresis impact within the same computational framework. This provides a basis for subsequent preamble speech retention length control and continuous acceptance speech segment generation. The preamble acceptance constraint value is dimensionless. The preamble retention compression term, link establishment suppression term, and hysteresis attenuation term are derived from the retrievable speech retention duration data in the circular speech buffer, the low-power wake-up trigger timestamp, the main processing power-on completion timestamp, the main clock lock timestamp, the speech recognition model loading completion timestamp, and the continuous speech takeover start timestamp, corresponding to the relevant basic data. Further calculations revealed that the relevant basic data had undergone standardization and normalization during the preprocessing stage, and participated in the calculation of the preamble constraint value in a dimensionless form during subsequent calculations. Based on this, after mapping using the arctangent function, inverse hyperbolic sine function, and exponential function, the output result of the preamble constraint value does not carry physical units. This result serves as the control basis for determining the preamble readback duration of the circular speech buffer, initiating dual-path mirror writing, and re-trunculating the splicing start point. This establishes a correspondence between the preamble constraint value and the circular speech buffer retention control, speech segment splicing control, and continuous speech takeover control. The specific calculation formula is as follows:

[0040] ;

[0041] In the formula, This represents the preamble takeover constraint value, used to characterize the combined constraint degree of preamble voice preservation and continuous takeover under the current link switching conditions; This indicates the write duration of the circular speech buffer, used to characterize the range of preamble speech samples that can be read back; This represents the low-power wake-up trigger timestamp, used to characterize the start time of the low-power monitoring link output wake-up determination; This indicates the timestamp when the main processing unit completes power-on, used to characterize the moment when the main processing link enters the startable state; This represents the master clock lock timestamp, used to indicate the moment when the master processing link clock is established; This represents the timestamp indicating the completion of speech recognition model loading, used to characterize the moment when speech recognition computing resources are fully loaded; This indicates the continuous voice takeover start timestamp, used to characterize the moment when the main processing link begins receiving continuous voice data.

[0042] In this embodiment, Table 1 is a data table of key parameters and constraint values ​​for preamble acceptance constraints. It details the circular buffer write time, link establishment time product, acceptance lag time difference, preamble retention compression term, link establishment suppression term, acceptance lag attenuation term, and the finally calculated preamble acceptance constraint value for different samples during the voice wake-up link switching process. This data is used to quantify the constraint capability of preamble voice acceptance under the hierarchical monitoring architecture. Specifically: Sample 1 has a circular buffer write time of 0.10, a link establishment time product of 0.01, an acceptance lag time difference of 0.05, a preamble retention compression term of 0.8329, a link establishment suppression term of 1.0100, an acceptance lag attenuation term of 0.9550, and a preamble acceptance constraint value of 0.7875; Sample 2 has a circular buffer write time of 0.10, a link establishment time product of 0.04, an acceptance lag time difference of 0.10, and a preamble retention... The compression term is 0.8329, the link establishment suppression term is 1.0398, the hysteresis attenuation term is 0.9139, and the leading hysteresis constraint value is 0.7322. For sample 3, the circular buffer write duration is 0.30, the link establishment time product is 0.01, the hysteresis time difference is 0.05, the leading retention compression term is 0.9151, the link establishment suppression term is 1.0100, the hysteresis attenuation term is 0.9624, and the leading hysteresis constraint value is 0.871. 7; Sample 4 has a ring buffer write time of 0.30, a link establishment time product of 0.04, a take-off lag time difference of 0.10, a leading retention compression term of 0.9151, a link establishment suppression term of 1.0398, a take-off lag attenuation term of 0.9262, and a leading take-off constraint value of 0.8156; Sample 5 has a ring buffer write time of 0.60, a link establishment time product of 0.01, a take-off lag time difference of 0.20, and a leading retention compression term of 0.9151, a link establishment suppression term of 1.0398, a take-off lag attenuation term of 0.9262, and a leading take-off constraint value of 0.8156. The compression term is 1.0122, the link establishment suppression term is 1.0100, the lag attenuation term is 0.8825, and the leading lag constraint value is 0.8849; the writing time of the ring buffer corresponding to sample 6 is 0.60, the link establishment time product is 0.08, the lag time difference is 0.30, the leading retention compression term is 1.0122, the link establishment suppression term is 1.0798, the lag attenuation term is 0.8353, and the leading lag constraint value is 0.7835.

[0043] Table 1 Key Parameters and Constraint Values ​​of Leading Bearing Constraints

[0044] Circular buffer write duration Link establishment time product Adapting to the time difference Pre-retained compression terms Link establishment suppression term Receive the hysteresis attenuation term Preceding constraint value J 0.10 0.01 0.05 0.8329 1.0100 0.9550 0.7875 0.10 0.04 0.10 0.8329 1.0398 0.9139 0.7322 0.30 0.01 0.05 0.9151 1.0100 0.9624 0.8717 0.30 0.04 0.10 0.9151 1.0398 0.9262 0.8156 0.60 0.01 0.20 1.0122 1.0100 0.8825 0.8849 0.60 0.08 0.30 1.0122 1.0798 0.8353 0.7835

[0045] like Figure 3The figure shows the density distribution and confidence elliptic plot of the leading acceptor constraint value. Combined with Table 1, it can be seen that the distribution of the leading acceptor constraint value with the circular buffer write duration and acceptor lag difference exhibits a clear pattern. Specifically, samples 3 and 5 have relatively high leading acceptor constraint values ​​of 0.8717 and 0.8849, respectively, corresponding to relatively large circular buffer write durations of 0.30 and 0.60, and moderate acceptor lag differences of 0.05 and 0.20. This indicates that the leading acceptor constraint capability is strongest under conditions of large buffer capacity and moderate switching latency. Sample 2 has the lowest leading acceptor constraint value of 0.7322, corresponding to a relatively small circular buffer write duration of 0.10 and a relatively large acceptor lag difference of 0.10, indicating that a small buffer and long switching latency severely weaken the leading acceptor capability. The density contour lines show that the sample points are mainly clustered in the region with a ring buffer duration of 0.10-0.30 seconds and a lag time difference of 0.05-0.10 seconds. The 68% confidence ellipse, drawn with a red dashed line, further outlines the typical distribution range of the parameter combination. Overall, this figure intuitively reveals that increasing the ring buffer write duration and controlling the lag time difference helps improve the leading lag constraint value, providing a quantitative basis for subsequent continuous speech reconstruction and boundary correction.

[0046] like Figure 4 The bar chart showing the comparison of preamble acceptance constraints is presented. Table 1 illustrates the significant differences in preamble acceptance constraints for each parameter combination. Sample 5, with its parameter combination of 0.60 for the ring buffer duration, 0.01 for the link time product, and 0.20 for the acceptance time difference, exhibits the highest preamble acceptance constraint value of 0.8849. This indicates that under the combination of a large ring buffer, extremely short link establishment time, and a moderate acceptance time difference, the preamble voice acceptance is most complete. Sample 3, with its parameter combination of 0.30 for the ring buffer duration, 0.01 for the link time product, and 0.05 for the acceptance time difference, has the second highest constraint value of 0.8717. Sample 2, with its parameter combination of 0.10 for the ring buffer duration, 0.04 for the link time product, and 0.10 for the acceptance time difference, has the lowest constraint value of only 0.7322. This reflects that when a small ring buffer, long link establishment time, and a relatively long acceptance time difference coexist, the risk of preamble voice loss is highest. Extending to 1.0 on the vertical axis clearly shows the gap between each sample constraint value and the ideal maximum value. Overall, this bar chart intuitively compares the preamble reception performance under different parameter configurations. It can be used to guide the selection of parameter combinations in actual systems that have a longer ring buffer duration, a smaller link establishment time product, and a moderate reception lag time difference, in order to reduce the probability of preamble voice loss during voice wake-up switching.

[0047] This implementation scheme achieves joint quantitative characterization of the preamble voice retention status, link establishment status, and continuous voice takeover status during the switch from low-power monitoring link to main processing link. This enables the preamble voice retention capability, link establishment time, and takeover lag to form comparable, calculable, and controllable constraint results under the same discriminative framework. It enhances the ability to identify risks of preamble voice truncation, takeover timing mismatch, and insufficient continuous takeover during the link switching phase. It provides a unified control basis for ring voice buffer retention control, preamble readback duration determination, dual-path mirror writing initiation, and splicing start point re-truncation, thereby improving the stability of the continuous takeover voice segment generation process and the integrity of preamble voice retention.

[0048] Specifically, the process of controlling the preamble length and generating consecutive speech segments based on the preamble constraint analysis results is as follows: Real-time comparison of the preamble constraint value and the preamble constraint threshold:

[0049] When the preceding continuation constraint value is greater than or equal to the preceding continuation constraint threshold, the length of the preceding speech already retained in the circular speech buffer is kept from expanding further. The preceding speech segment is read back from the circular speech buffer according to the preceding readback duration n before the low-power wake-up trigger timestamp. The preceding readback duration n is determined by the time difference between the low-power wake-up trigger timestamp and the continuous speech takeover start timestamp, the writing duration of the circular speech buffer, and the starting position of the wake-up candidate segment in the original speech waveform data. The value range is the time interval between the starting position of the wake-up candidate segment and the starting position of the circular speech buffer that can be read back, and does not exceed the writing duration of the circular speech buffer. The preceding speech segment read back and the newly sampled speech segment after the continuous speech takeover start timestamp are then subjected to double-end splicing. During splicing, a sample rearrangement method based on timestamp alignment is used to eliminate sampling boundary misalignment, and then an overlapping addition process based on the energy consistency of adjacent speech analysis intervals in the overlapping area is used to eliminate splicing abrupt changes, forming a continuous continuation speech segment.

[0050] When the leading acceptance constraint value is less than the leading acceptance constraint threshold, the retention time of the circular voice buffer is extended to maintain the synchronous retention of voice data within the acceptance retention time interval between the low-power monitoring link and the main processing link. The acceptance retention time interval is determined by the time difference between the low-power wake-up trigger timestamp, the master clock lock timestamp, and the continuous voice takeover start timestamp. Its starting point is the leading retention start position corresponding to the low-power wake-up trigger timestamp, and its ending point is the end position of the first stable vocal interval after the continuous voice takeover start timestamp. The duration of the acceptance retention time interval does not exceed the writing time of the circular voice buffer. During the acceptance retention time interval, dual-path mirror writing is performed on the leading voice segment and the newly sampled voice segment, so that the leading voice segment and the newly sampled voice segment are respectively retained in the corresponding retention interval of the circular voice buffer and the corresponding takeover interval of the main processing link. Then, the splicing starting point is re-trimmed according to the low-power wake-up trigger timestamp, the master clock lock timestamp, and the continuous voice takeover start timestamp, and the continuous voice takeover start time is used. The start timestamp is used as the truncation reference. Truncation position offset correction is performed by combining the low-power wake-up trigger timestamp and the master clock lock timestamp. The truncated speech segments are re-weighted and spliced ​​in the overlapping area with continuously varying weighting coefficients to fill speech gaps and prevent the wake-up word and control statement from being cut off. The overlapping area is the continuous overlapping sampling interval formed by the end of the preceding speech segment and the beginning of the newly sampled speech segment after the re-truncation and splicing start point, with an overlap length of 5 to 40 milliseconds. The weighting coefficients in the overlapping area are determined by the relative positions of the sampling points in the overlapping area. The weighting coefficients corresponding to the preceding speech segment and the newly sampled speech segment are both continuously varying values ​​between 0 and 1, and the sum of the weighting coefficients of the two at the same sampling position is 1. This causes the weighting coefficient corresponding to the preceding speech segment to continuously decrease along the time progression direction, and the weighting coefficient corresponding to the newly sampled speech segment to continuously increase along the time progression direction. After compensation, a continuous speech segment is generated, and then continuous speech reconstruction and boundary correction processing is performed.

[0051] This implementation scheme achieves hierarchical control of the ring speech buffer retention strategy, preamble readback strategy, dual-path mirror writing strategy, and splicing compensation strategy based on the preamble continuation constraint value. This enables the preamble speech segment and the newly sampled speech segment to be processed according to the difference in continuation status during the switch from the low-power monitoring link to the main processing link. This reduces the problems of missing wake-up words, truncated control statements, and abrupt splicing boundaries caused by insufficient preamble retention, splicing start point offset, and speech continuation gaps during the link switching stage. It enhances the speech continuity, boundary stability, and temporal sequence integrity during the generation of continuous continuation speech segments, providing a more complete and stable speech input foundation for subsequent continuous speech reconstruction and boundary correction processing.

[0052] Specifically, the process of restoring and correcting the boundary offset, spectral continuity, and silence breakage of consecutive speech segments is as follows: The wake-up tail-end rollback amount is calculated from the log-Mel spectrum sequence of the consecutive speech segments using a wake-up word template dynamic time warping algorithm. The frame-level path is aligned between the log-Mel spectrum sequence of the consecutive speech segments and the wake-up word template spectrum sequence. The end point of the tail-end offset path is determined based on the principle of minimizing cumulative mismatch cost. The wake-up tail-end rollback amount is then calculated from the displacement of the path end point relative to the template standard end point. The log-Mel spectrum sequences on both sides of the connection position formed when the leading speech segment and the newly sampled speech segment after the start timestamp of the consecutive speech takeover in the consecutive speech segments are processed using frequency... The cosine similarity of the spectrum is used to calculate the spectral continuity of the first segment. The cosine of the angle between the two spectral vectors formed by the two adjacent spectral frames at the connection position is calculated to characterize the consistency of the frequency band energy distribution before and after splicing. For the original speech waveform data and speech sampling timestamp sequence of continuous speech segments, the energy statistics of adjacent speech analysis intervals and the joint detection of zero crossing rate are used to identify silent discontinuities with continuously decreasing energy and synchronously changing zero crossing rate. Then, the density of broken silence is calculated based on the frequency of occurrence and the proportion of duration of silent discontinuities within a unit of time. Adjacent speech analysis intervals are divided into continuous equal-length analysis units according to the speech sampling timestamp sequence. The energy integral value and the number of zero crossings of each analysis unit are jointly judged. Then, the density statistics are performed on the analysis units that meet the silent discontinuity characteristics.

[0053] Calculate the integral value of the sum of one divided by one and the square of the integral variable within the interval from zero to the wake-up tail-end rollback. Take the negative exponent of the integral value to obtain the boundary offset suppression term. Compress the growth rate of the wake-up tail-end rollback within a large offset interval through bounded integral mapping, and further enhance the attenuation effect of the tail-end offset on the reliable results through negative exponential mapping. Calculate the error function value of the segment-first spectral connectivity, add one, and obtain the spectral connectivity enhancement term. Perform a smooth nonlinear expansion of the segment-first spectral connectivity through the error function to ensure that the spectral connectivity difference remains monotonically separable within the continuous interval. The addition of one ensures that the enhancement term is always positive; the sum of one and the fracture silence density is calculated, and then the natural logarithm is taken and one is added to obtain the silence fracture suppression term; the numerical span of the fracture silence density is compressed by the natural logarithm to reduce the amplification of the result when the number of silence discontinuities increases sharply, and the addition of one is used to maintain the stability of the denominator term; the boundary offset suppression term is multiplied by the spectral connection enhancement term and then divided by the silence fracture suppression term to obtain the boundary restoration reliability value, which is used to comprehensively characterize the combined influence of tail offset, splicing connection and silence fracture on the reliability of boundary restoration within the same calculation framework. The boundary restoration reliability value is a dimensionless quantity. The wake-up tail rollback amount, segment-first spectral continuity, and broken silence density are further calculated from the relevant basic data corresponding to the continuous speech segments, log-Mel spectrum sequence, original speech waveform data, and speech sampling timestamp sequence. Among them, the relevant basic data has been standardized and normalized in the preprocessing stage, and participates in the calculation of boundary offset suppression term, spectral continuity enhancement term, silence breakage suppression term, and boundary restoration reliability value in a dimensionless form in the subsequent calculation process. On this basis, after integral mapping, error function mapping, natural logarithm mapping, and multiplication and division combination processing, the boundary restoration reliability value output result does not carry physical units. The boundary restoration reliability value is constructed to address the boundary drift problem at the splicing point of the preceding speech segment and the newly sampled speech segment during the switching between the low-power monitoring link and the main processing link. It jointly represents the wake-up tail misalignment, splicing silence breakage, and connection interval spectral mismatch caused by link switching, and serves as the linkage control basis for connection position forward rollback, silence interval deletion or merging, and connection correction interval elastic resampling. The specific calculation formula is as follows:

[0054] ;

[0055] In the formula, The boundary restoration confidence value is used to characterize the degree to which consecutive speech segments retain the recognizability of speech boundaries after boundary restoration. This indicates the amount of wake-up word pullback, used to characterize the degree of offset of the wake-up word's tail relative to the template's standard endpoint; This indicates the degree of spectral coherence at the beginning of a segment, used to characterize the consistency of the spectral distribution on both sides of the connection point; It represents the density of broken silences, used to characterize the degree to which silence breaks the continuity of speech; This represents the error function mapping value corresponding to the spectral connectivity at the beginning of a segment, used to characterize the enhancement result of the spectral connectivity at the beginning of a segment after nonlinear smoothing mapping.

[0056] This implementation scheme achieves joint identification and unified quantitative characterization of abnormal splicing boundaries of continuously received speech segments during the switching process between the low-power monitoring link and the main processing link. This enables the degree of wake-up tail misalignment, the spectral connection status of the connection position, and the degree of silence break interference to form calculable, comparable, and controllable reliable results under the same discriminative framework. It enhances the ability to identify boundary drift, spectral mismatch, and speech continuity disruption problems in link switching scenarios. It provides a unified control basis for forward repositioning of connection positions, deletion or merging of silence interruption zones, and elastic resampling of connection correction intervals. This improves the accuracy of the process of determining the end position of the wake-up word and the start position of the control statement, the stability of speech boundary restoration, and the completeness of subsequent password recognition input.

[0057] Specifically, the process of correcting the end position of the wake-up word, correcting the start position of the control statement, and reconstructing continuous speech is as follows: Figure 5 The diagram shows the flowchart for boundary restoration confidence value determination and speech segment processing, which compares the boundary restoration confidence value and the boundary restoration confidence threshold in real time.

[0058] When the boundary restoration confidence value is greater than or equal to the boundary restoration confidence threshold, the path endpoint obtained by performing wake-up word template dynamic time warping matching on the log-Mel spectrum sequence of the continuous speech segment is used as the wake-up tail truncation position. The continuous speech segment is divided into wake-up word segments and control statement segments according to the wake-up tail truncation position. The silence boundary is redefined on the segmented control statement segments. The silence boundary is determined based on the energy integral value, zero-crossing count, and continuous silence duration of adjacent speech analysis intervals. The starting point of the first speech analysis interval that satisfies the initial phonation energy surge and the zero-crossing count change synchronously is determined as the first phonation point. A preset number of initial retention samples are retained before the first phonation point to prevent the initial consonant of the first word from being truncated again. Then, the reconstructed complete control statement segment is used for password recognition.

[0059] When the boundary restoration confidence value is less than the boundary restoration confidence threshold, restoration processing is performed on the continuous voice segments: the connection position formed by the preamble voice segment and the newly sampled voice segment after the start timestamp of the continuous voice takeover is moved forward during the double-end splicing process according to the wake-up tail segment back amount; the silence intervals in the adjacent voice analysis intervals on both sides of the connection position are deleted or merged according to the broken silence density; the connection correction interval is elastically resampled in the spectral connection interval on both sides of the connection position according to the segment start spectral coherence. The connection correction interval is the continuous voice analysis interval within a preset length range on both sides of the connection position. The stretching ratio and compression ratio of the elastic resampling are limited to the preset ratio range, and the sampling point position after resampling is reconstructed by continuous interpolation so that the syllable transition on both sides of the connection position converges simultaneously in the time domain and frequency domain; after the processing is completed, the corrected wake-up word end position and control statement start position are determined, and then password recognition is performed.

[0060] This implementation scheme achieves hierarchical processing of wake-up word tail segment truncation, control statement start boundary preservation, and splicing anomaly restoration strategies based on boundary restoration reliability values. This enables the stable segmentation of wake-up word fragments and control statement fragments when the boundary restoration state meets the requirements. When the boundary restoration state is insufficient, it can perform directional correction for connection position offset, silence breakage, and spectrum connection mismatch. This reduces the interference of wake-up tail segment truncation, control statement first segment missing, and splicing transition instability on the password recognition results in link switching scenarios. It enhances the accuracy of boundary determination, syllable transition continuity, and voice input integrity in the control statement fragment reconstruction process, providing a more reliable voice foundation for subsequent password recognition and control command generation.

[0061] Specifically, the process of performing execution closure analysis based on home appliance control execution data is as follows: For complete control statement fragments, a posterior probability sequence of commands is obtained through a speech recognition model. Then, the posterior setpoints of the commands are obtained by mapping using a Gaussian cumulative distribution function. The speech recognition model outputs the posterior probability distribution of each candidate command category for the complete control statement fragment. The posterior setpoints of the commands are used to characterize the degree to which the posterior probability quality is concentrated towards the target command category; their value increases as the proportion of the posterior probability corresponding to the target command category increases. Then, a monotonic mapping of the concentration of the posterior probability distribution is performed using a Gaussian cumulative distribution function to obtain a stable expression of the posterior setpoints of the commands. The mean and variance in the Gaussian cumulative distribution function are determined by the statistical results of the posterior probability distribution output by the speech recognition model for the candidate command categories. The complete control statement fragment is decoded by the speech recognition model and identified using device word extraction rules to obtain the device referential text. Then, the device referential text is compared with the home appliance type data. The device word matching degree is calculated using the edit distance normalization algorithm. The device word extraction rule is used to locate the text segment corresponding to the device name from the decoding result. The edit distance normalization algorithm is used to calculate the character-level difference ratio between the device reference text and the home appliance type data. For complete control statement fragments and home appliance executable instruction set data, the instruction slot missing degree is calculated using the intent slot filling algorithm. The intent slot filling algorithm maps slots according to the action expression, object expression, and parameter expression in the complete control statement fragment, and counts the proportion of unfilled target slots. The instruction slot missing degree is the ratio of the number of unfilled target slots to the total number of target slots required for the current home appliance executable instruction. The execution receipt closure time difference is calculated by the difference between the home appliance control instruction issuance timestamp and the home appliance execution receipt timestamp. The execution receipt closure time difference is used to characterize the timing closure distance between the issuance of the control instruction and the return of the execution feedback.

[0062] The command-device coordination term is obtained by multiplying the command posterior set value by the sum of the device word matching degree and one. This product structure creates a joint enhancement relationship between the command recognition set degree and the device correspondence degree. Adding one prevents compression imbalance in the product term when the device word matching degree approaches zero. The execution damping term is obtained by calculating the square root of the sum of the square of the instruction slot missing degree and the square of the execution acknowledgment closure time difference, and then adding one. A joint damping norm is constructed by taking the square root of the sum of the squares, ensuring that the instruction slot missing degree and the execution acknowledgment closure time difference participate in the suppression operation in a unified manner. Adding one ensures that the denominator remains positive and maintains computational stability. The command-device coordination term is divided by the execution damping term, then one is added, and the natural logarithm is taken to obtain the password execution confidence value. This ratio structure highlights the enhancing effect of the command-device coordination term on the confidence result, and the natural logarithm compresses the growth rate within a large numerical range, keeping the output result smooth and bounded. The password execution reliability value is a dimensionless quantity. The command posterior set value, device word matching degree, command slot missing degree, and execution receipt closure time difference are obtained from relevant basic data through subsequent calculations. The relevant basic data has undergone standardization and normalization in the preprocessing stage and participates in the calculation of command-device coordination terms, execution damping terms, and password execution reliability value in a dimensionless form. Based on this, after summation of squares, square root, addition of one, and natural logarithm mapping, the password execution reliability value output does not carry physical units. The specific calculation formula is as follows:

[0063] ;

[0064] In the formula, This indicates the password execution confidence value, which characterizes the overall executability of a complete control statement fragment under the conditions of device correspondence, slot integrity, and execution closure. This represents the set value of the command posterior, used to characterize the degree of concentration of the command posterior probability sequence in the candidate command space; This indicates the degree of matching between device terms, which characterizes the consistency between the device reference text and the home appliance type data. Indicates the instruction slot missing degree, used to characterize the degree of slot missing of complete control statement fragments under the constraints of the executable instruction set data of home appliances; This indicates the execution receipt closure time difference, used to characterize the degree of timing closure between the issuance of home appliance control commands and the return of execution receipts.

[0065] This implementation scheme achieves joint quantitative characterization of complete control statement fragments in terms of command recognition concentration, device correspondence consistency, slot filling completeness, and execution closure timing. This enables the executability of password content, device matching effectiveness, and control closure reliability to form comparable, calculable, and verifiable credible results under the same discriminative framework. It reduces the interference of device misassignment, missing action parameters, and execution feedback lag on the accuracy of control decisions, and enhances the ability to identify abnormal passwords, screen ambiguous devices, and suppress incomplete control outputs in error wake-up scenarios. It provides a unified discriminative basis for subsequent restricted word list error correction decoding, slot completion re-parsing, generation of unique home appliance control instructions, and control execution closure verification.

[0066] Specifically, the process of confirming control semantics, issuing home appliance control commands, and verifying execution results in a closed loop based on the results of the closed-loop analysis is as follows: Real-time comparison of the password execution trust value and the password execution trust threshold:

[0067] When the password execution trust value is less than the password execution trust threshold, appliance control commands are not directly issued. Instead, a restricted vocabulary error correction decoding and slot completion re-parsing are performed on the complete control statement fragment. The restricted vocabulary consists of appliance executable instruction set data, appliance type data, and a set of synonym expressions for device names. For cases where the device reference text cannot uniquely correspond to a single appliance type, nearest neighbor matching is re-performed in the appliance type data. Nearest neighbor matching uses an edit distance normalization algorithm to calculate the degree of difference between the device reference text and each appliance type data. The re-matched device set then constrains the decoding space, limiting the re-decoding process to the specified range. The candidate control items corresponding to the device set are filtered and converged. If the slot filling result corresponding to the missing instruction slot does not cover the target action slot and target parameter slot in the current home appliance executable instruction set, the missing action and parameter are filled in according to the home appliance executable instruction set data. After re-decoding, the password execution confidence value is recalculated. If the recalculated password execution confidence value is greater than or equal to the password execution confidence threshold, the control command is issued. If the recalculated password execution confidence value is still less than the password execution confidence threshold, only a local voice confirmation prompt is output, and no control is issued, thereby avoiding erroneous control after erroneous wake-up.

[0068] When the password execution trust value is greater than or equal to the password execution trust threshold, candidate control items consistent with the current device are first screened from the executable instruction set data of the home appliance based on the home appliance type data. Then, a restricted vocabulary secondary decoding is performed on the complete control statement fragment to ensure that the decoding result converges only within the vocabulary range corresponding to the candidate control item. After the secondary decoding is completed, a unique home appliance control instruction is generated and sent to the target device corresponding to the home appliance identifier through the control bus. After the instruction is sent, the home appliance execution receipt status data and the home appliance execution receipt timestamp are continuously read. The control instruction and execution receipt are matched one-to-one to confirm whether the current round of control is closed. When the verification is consistent, the final execution result is output.

[0069] This implementation scheme achieves hierarchical control and closure verification of complete control statement fragments based on the credibility value of the command execution. This enables voice control requests with ambiguous device designations, missing slots, and insufficient execution credibility to complete the restricted word list error correction decoding, nearest neighbor rematching, and slot completion processes before formal issuance. It also enables control requests that meet the execution conditions to generate unique home appliance control instructions under the constraints of candidate control items and complete the execution receipt closure confirmation. This reduces the impact of false wake-up, mismatched devices, missing action parameters, and abnormal execution feedback on the accuracy of control results, and enhances the determinism of the home appliance control instruction generation process, the effectiveness of control issuance, and the verifiability of execution results.

[0070] like Figure 2As shown, the second aspect of the present invention provides a voice-controlled smart home appliance wake-up system, comprising: a data acquisition and preprocessing module, used to acquire observation data of the voice wake-up process and obtain home appliance control execution data, and to perform time-series unification, quality ordering, and standardized expression on the voice wake-up process observation data and home appliance control execution data; to achieve unified organization and standardized processing of multi-source asynchronous data in the low-power monitoring link, main processing link, and control execution link, thereby improving the data consistency and comparability of subsequent cross-stage state discrimination and joint calculation; and a preamble constraint discrimination module, used to perform preamble constraint analysis based on link switching observation data, and to control the preamble voice retention length and generate continuous successive voice segments based on the preamble constraint analysis results; to achieve joint constraint control of the preamble voice retention state, link establishment state, and continuous takeover state during link switching, thereby improving the preservation of the preamble voice. The system ensures the integrity and stability of continuously generated speech segments. A continuous speech reconstruction and boundary correction module is used to analyze and correct boundary offsets, spectral connections, and silence breaks in continuously generated speech segments, and to correct the end position of wake-up words, the start position of control statements, and the continuous speech reconstruction. This enables joint correction of wake-up tail misalignment, splicing boundary mismatch, and silence breaks in link switching scenarios, improving the accuracy of control statement boundary determination and the integrity of voice input. A password recognition and execution closure module is used to perform execution closure analysis based on appliance control execution data, and to confirm control semantics, issue appliance control commands, and verify execution results based on the analysis results. This enables unified verification of control semantic validity, device matching status, slot integrity status, and execution closure status, improving the accuracy of appliance control command generation, the effectiveness of control issuance, and the verifiability of execution results.

[0071] This implementation scheme achieves full-process collaborative processing of voice wake-up, voice reception, boundary restoration, command recognition, and control execution closure during the switch from the low-power monitoring link to the main processing link. This enables multi-source asynchronous observation data, preamble voice retention status, continuous voice boundary status, and home appliance control execution status to form continuous connection, hierarchical discrimination, and closed-loop verification under a unified framework. This reduces the impact of preamble voice segment loss, splicing boundary drift, incomplete control statement reconstruction, and execution result mismatch on the accuracy of home appliance voice control under the existing hierarchical monitoring and time-division wake-up architecture. It improves the stability of continuous reception voice segment generation, the completeness of control statement fragment reconstruction, the accuracy of home appliance control command issuance, and the closed-loop confirmability of execution results.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0073] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for intelligent wake-up of home appliances based on voice control, characterized in that, Includes the following steps: S1, collect observation data of the voice wake-up process and obtain home appliance control execution data, and perform time-series unification, quality-ordering and standardized expression of the voice wake-up process observation data and home appliance control execution data; S2, based on link switching observation data, performs preamble continuation constraint analysis, and controls the preamble speech retention length and generates continuous continuation speech segments based on the preamble continuation constraint analysis results; S3 performs restoration and correction analysis on the boundary offset, spectral connection relationship and silence break state of continuous speech segments, and performs correction of the end position of wake-up word, correction of the start position of control statement and reconstruction of continuous speech. S4 combines the execution data of home appliance control to perform execution closure analysis, and performs control semantic confirmation, home appliance control command issuance and execution result closure verification based on the execution closure analysis results.

2. The method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process of collecting observation data during the voice wake-up process and obtaining home appliance control execution data is as follows: In the home appliance voice wake-up terminal, observation data of the voice wake-up process and data on home appliance control execution are collected. The data collected includes: continuously collecting raw indoor voice waveform data through analog-to-digital conversion sampling circuit and microphone array, and synchronously recording the voice sampling timestamp sequence; real-time reading of the replayable voice retention time data, low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, voice recognition model loading completion timestamp, and continuous voice takeover start timestamp in the circular voice buffer; data on home appliance control execution is obtained by: obtaining home appliance identification, home appliance type data, and home appliance executable instruction set data through the local device registry and local control protocol table; and reading the home appliance control instruction issuance timestamp, home appliance execution receipt timestamp, and home appliance execution receipt status data through the control bus log and execution receipt log.

3. The method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process of unifying the timing, sorting the quality, and standardizing the expression of the observation data of the voice wake-up process and the home appliance control execution data is as follows: The voice wake-up process observation data and home appliance control execution data are aligned with unified timestamps and duplicate records are removed; isolated spikes are suppressed in the original voice waveform data by detecting abrupt amplitude changes between adjacent sampling points. The system performs missing record completion and protocol consistency verification on the home appliance control execution data; it extracts the time-spectrum sequence from the original speech waveform data through short-time Fourier transform processing, and extracts the log-Mel spectrum sequence from the time-spectrum sequence through log-Mel spectrum extraction processing; it standardizes the log-Mel spectrum sequence, voice wake-up process observation data, and home appliance control execution data through the Z-score normalization algorithm; and it normalizes the low-power wake-up trigger timestamp, main processing power-on completion timestamp, main clock lock timestamp, speech recognition model loading completion timestamp, continuous voice takeover start timestamp, home appliance control command issuance timestamp, and home appliance execution receipt timestamp through the maximum and minimum value normalization algorithm.

4. The method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process of performing the leading constraint analysis based on link switching observation data is as follows: Acquire the data on the duration of replayable voice retention, the low-power wake-up trigger timestamp, the main processing power-on completion timestamp, the main clock lock timestamp, the voice recognition model loading completion timestamp, and the continuous voice takeover start timestamp; Calculate the sum of the writing duration of the first and the circular speech buffer, take the arctangent of the result, and obtain the preamble-preserving compression term. Calculate the difference between the speech recognition model loading completion time stamp and the master clock locking time stamp, calculate the difference between the master clock locking time stamp and the main processing power-on completion time stamp, calculate the product of the two differences, take the inverse hyperbolic sine and add one to obtain the link establishment suppression term; The difference between the continuous voice takeover start timestamp and the low-power wake-up trigger timestamp is calculated, divided by the sum of the writing duration of the one-to-one and circular voice buffers, and then the negative exponent is taken to obtain the takeover hysteresis attenuation term; the preceding retention compression term is divided by the link establishment suppression term, and then multiplied by the takeover hysteresis attenuation term to obtain the preceding takeover constraint value.

5. The method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process of controlling the preamble length and generating continuous speech segments based on the preamble constraint analysis results is as follows: Real-time comparison of leading constraint value and leading constraint threshold: When the preceding acceptance constraint value is greater than or equal to the preceding acceptance constraint threshold, the length of the preceding speech already retained in the circular speech buffer is maintained. The preceding speech segment is read back from the circular speech buffer according to the preceding readback duration n before the low-power wake-up trigger timestamp. The readback preceding speech segment is then combined with the newly sampled speech segment after the continuous speech takeover start timestamp to form a continuous acceptance speech segment. When the preceding acceptance constraint value is less than the preceding acceptance constraint threshold, the retention time of the circular voice buffer is extended to maintain the synchronous retention of voice data within the acceptance retention time interval between the low-power monitoring link and the main processing link. During the acceptance retention time interval, dual-path mirror writing is performed on the preceding voice segment and the newly sampled voice segment, so that the preceding voice segment and the newly sampled voice segment are respectively retained in the corresponding retention interval of the circular voice buffer and the corresponding takeover interval of the main processing link. Then, the splicing starting point is re-trimmed based on the low-power wake-up trigger timestamp, the main clock lock timestamp, and the continuous voice takeover start timestamp. The truncation position offset correction is performed with the position corresponding to the continuous voice takeover start timestamp as the truncation reference, combined with the low-power wake-up trigger timestamp and the main clock lock timestamp. The method of re-weighting and splicing the overlapping areas of the extracted speech segments using continuously varying weighting coefficients is used to fill the gaps in speech continuity. After compensation, continuous speech segments are generated, and then continuous speech reconstruction and boundary correction processing is performed.

6. The method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process for restoring and correcting the boundary offset, spectral continuity, and silence breakage of consecutive speech segments is as follows: The wake-up tail rollback amount is calculated by using the wake-up word template dynamic time warping algorithm for the log-Mel spectrum sequence of consecutive speech segments; the segment-first spectrum coherence degree is calculated by using the spectrum cosine similarity to obtain the log-Mel spectrum sequence on both sides of the connection position formed by the preceding speech segment and the newly sampled speech segment in the consecutive speech segment during the double-end splicing process; the silence discontinuity region with continuously decreasing energy and synchronously changing zero-crossing rate is identified by using energy statistics and zero-crossing rate joint detection of adjacent speech analysis intervals for the original speech waveform data and speech sampling timestamp sequence of consecutive speech segments, and then the broken silence density is calculated based on the frequency of occurrence and the proportion of duration of the silence discontinuity region within a unit duration. Calculate the integral value of the sum of 1 divided by 1 and the square of the integral variable within the interval from zero to the wake-up tail segment. Take the negative exponent of the integral value to obtain the boundary offset suppression term. Calculate the error function value of the spectral connectivity at the beginning of the segment, and add 1 to obtain the spectral connectivity enhancement term. Calculate the sum of 1 and the break silence density, take the natural logarithm and add 1 to obtain the silence break suppression term. Multiply the boundary offset suppression term by the spectral connectivity enhancement term, and then divide by the silence break suppression term to obtain the boundary restoration confidence value.

7. The method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process of correcting the end position of the wake-up word, correcting the start position of the control statement, and reconstructing continuous speech is as follows: Real-time comparison of boundary restoration confidence value and boundary restoration confidence threshold: When the boundary restoration confidence value is greater than or equal to the boundary restoration confidence threshold, the path endpoint obtained by performing wake-word template dynamic time warping matching on the log-Mel spectrum sequence of the continuous speech segment is used as the wake-word tail truncation position. The continuous speech segment is divided into wake-word segments and control statement segments according to the wake-word tail truncation position. The silence boundary is redefined on the segmented control statement segments, and the initial retained sample points before the first utterance point are retained. Then, the reconstructed complete control statement segments are used for password recognition. When the boundary restoration confidence value is less than the boundary restoration confidence threshold, restoration processing is performed on the continuous speech segments: the connection position formed during the double-end splicing process of the newly sampled speech segment is pushed forward according to the wake-up tail segment back amount; the silence intervals in the adjacent speech analysis intervals on both sides of the connection position are deleted or merged according to the break silence density; the connection correction interval is elastically resampled in the spectral connection interval on both sides of the connection position according to the segment start spectral connection degree; after the processing is completed, the corrected wake-up word end position and control statement start position are determined, and then password recognition is performed.

8. The method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process of performing execution closure analysis by combining home appliance control execution data is as follows: The complete control statement fragment is used to obtain the command posterior probability sequence through a speech recognition model, and then the command posterior set value is obtained by mapping through the Gaussian cumulative distribution function. The complete control statement fragment is decoded by a speech recognition model and combined with the device word extraction rules to obtain the device reference text. The device reference text and the home appliance type data are then used to calculate the device word matching degree through the edit distance normalization algorithm. For complete control statement fragments and executable instruction set data of home appliances, the instruction slot missing degree is calculated by the intent slot filling algorithm; the execution receipt closing time difference is calculated by the difference between the timestamp of the home appliance control instruction issuance and the timestamp of the home appliance execution receipt. The command-device coordination term is obtained by multiplying the value in the command posterior set with the sum of the device word matching degree and the value of 1. Calculate the square root of the sum of the square of the instruction slot missingness and the square of the execution receipt closing time difference, and add one to obtain the execution damping term; Divide the command device coordination term by the execution damping term, add one, and take the natural logarithm to obtain the password execution trust value.

9. A method for intelligent wake-up of home appliances based on voice control according to claim 1, characterized in that: The specific process of confirming control semantics, issuing home appliance control commands, and verifying execution results in a closed loop based on the results of the closed-loop analysis is as follows: Real-time comparison of password execution trust value and password execution trust threshold: When the password execution trust value is less than the password execution trust threshold, the complete control statement fragment is subjected to restricted vocabulary error correction decoding and slot completion re-parsing; for cases where the device reference text cannot be uniquely mapped to a single home appliance type, nearest neighbor matching is re-executed in the home appliance type data, and the decoding space is constrained by the re-matched device set; for cases where the slot filling result corresponding to the instruction slot missing degree does not cover the target action slot and target parameter slot in the current home appliance executable instruction set, the missing action and parameter are filled in according to the home appliance executable instruction set data. After re-decoding, the password execution trust value is recalculated. If the password execution trust value is still less than the password execution trust threshold, only a local voice confirmation prompt is output, and control is not issued. When the password execution trust value is greater than or equal to the password execution trust threshold, candidate control items consistent with the current device are screened from the executable instruction set data of the home appliance based on the home appliance type data. The complete control statement fragment is then subjected to secondary decoding using a restricted word list. After secondary decoding, a unique home appliance control instruction is generated and sent to the target device corresponding to the home appliance identifier via the control bus. After the instruction is sent, the home appliance execution receipt status data and the home appliance execution receipt timestamp are continuously read. The control instruction and execution receipt are matched one-to-one to confirm whether the current round of control is closed. When the verification is consistent, the final execution result is output.

10. A voice-controlled smart home appliance wake-up system, employing the voice-controlled smart home appliance wake-up method according to any one of claims 1-9, comprising: The acquisition and preprocessing module is used to acquire observation data of the voice wake-up process and obtain home appliance control execution data, and to perform time-series unification, quality-ordering and standardized expression on the voice wake-up process observation data and home appliance control execution data. The preamble constraint discrimination module is used to perform preamble constraint analysis based on link switching observation data, and to control the preamble speech retention length and generate continuous speech segments based on the preamble constraint analysis results; The continuous speech reconstruction and boundary correction module is used to restore and correct the boundary offset, spectral connection relationship and silence break state of continuous speech segments, and to perform wake-up word end position correction, control statement start position correction and continuous speech reconstruction. The password recognition and execution closure module is used to perform execution closure analysis by combining home appliance control execution data, and to perform control semantic confirmation, home appliance control command issuance and execution result closed-loop verification based on the execution closure analysis results.

Images