KTV song requesting system and method for requesting songs based on microphone voice instruction analysis

Abstract

The invention relates to the technical field of intelligent voice recognition, and discloses a KTV song requesting system and method for requesting songs based on microphone voice instruction analysis, and the system comprises a wireless microphone, a microphone receiver and a song requesting set top box. The wireless microphone is used for distributing a human voice signal into a sound amplification signal flow and an instruction signal flow which are parallel in real time during a key triggering period; the microphone receiver encapsulates the instruction signal flow into a data frame with a channel state identifier and audio data, and is connected to the song requesting set top box through a special data transmission line; the song requesting set top box comprises a synchronous sampling module and an instruction analysis engine, the synchronous sampling module is used for synchronously sampling the echo reference signal according to the channel state identifier, and the instruction analysis engine performs adaptive echo cancellation processing and semantic recognition on the audio data based on the echo reference signal so as to execute song requesting operation. The method is applied to the system. Interference such as background music can be effectively eliminated, and the voice instruction recognition accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent speech recognition technology, specifically a KTV song selection system and method based on microphone voice command parsing for song selection. Background Technology

[0002] With the development of voice control technology, intelligent voice assistants have been widely used in various consumer electronics and entertainment devices. In KTV settings, users increasingly demand to perform operations such as song selection and skipping directly by voice. However, applying this technology to the KTV environment faces unique and significant technical challenges.

[0003] Traditional voice assistants (such as smart speakers) typically employ a "hot word wake-up + continuous listening" working mode. This mode reveals significant drawbacks in KTV scenarios: First, the continuously playing high-volume background music and the mixed chatter and singing of multiple people in the room create strong environmental noise and interference echoes, easily drowning out the user's voice commands and causing a sharp drop in recognition rate. Second, this open continuous listening mode makes it impossible for the system to distinguish between the user's song selection command and irrelevant background dialogue or lyrics, easily leading to false wake-ups and misrecognitions, interfering with normal singing. Furthermore, the interaction method of requiring a fixed wake word (such as "Xiao Ai Tongxue") before each operation is cumbersome and unnatural in the KTV setting where frequent adjustments are needed, disrupting the continuity and immersion of the singing.

[0004] While existing technologies attempt to reduce noise and cancel echoes through software algorithms, in environments like KTVs with complex sound sources and strong signal coupling, the effectiveness of backend algorithms alone is limited and cannot fundamentally solve the problem of severe contamination of command signals during the input stage. Therefore, there is an urgent need for a KTV song selection technology solution that can innovate at the system level and ensure accurate, private, and convenient voice command interaction even in environments with strong interference. Summary of the Invention

[0005] The purpose of this application is to provide a KTV song selection system and method based on microphone voice command parsing for song selection, so as to solve the technical problems mentioned in the background art above.

[0006] To achieve the above objectives, this application discloses the following technical solutions: In a first aspect, this application discloses a KTV song selection system based on microphone voice command parsing for song selection, including: A wireless microphone includes a voice command button and a signal distribution circuit; the signal distribution circuit is configured to distribute the acquired human voice signal into a parallel amplified signal stream and a command signal stream in real time during the period when the voice command button is triggered. A microphone receiver is connected to the wireless microphone via a radio frequency link and is configured to encapsulate the received command signal stream into a data frame with channel status identifier and audio data via a data communication interface, and to connect to an audio amplifier via an audio output interface to output the amplified signal stream. The karaoke set-top box is connected to the data communication interface of the microphone receiver via a dedicated data transmission line. It includes an instruction parsing engine and a synchronous sampling module. The synchronous sampling module is configured to synchronously sample the background music digital signal output by the karaoke set-top box itself within an instruction parsing window determined based on the channel status identifier parsed from the data frame, using it as an echo reference signal. The instruction parsing engine is configured to perform adaptive echo cancellation processing on the audio data parsed from the data frame based on the echo reference signal, and to perform semantic recognition on the processed audio data to execute the corresponding song selection operation.

[0007] Optionally, the wireless microphone further includes a mute control module controlled by the voice command button; the mute control module is configured to attenuate or block the amplified signal stream when the voice command button is pressed, so that the human voice signal is transmitted to the karaoke set-top box only through the command signal stream.

[0008] Optionally, the adaptive echo cancellation process includes: Obtain the echo reference signal vector corresponding to the echo reference signal. ; The echo reference signal vector Input to a weighted vector An adaptive filter, wherein the weight vector Used for the echo reference signal vector Weighting is applied to simulate the echo path and generate an echo estimation signal; The residual signal is obtained by subtracting the echo estimation signal from the audio data of the data frame. , and serve as the target speech signal after echo cancellation; According to the residual signal With the echo reference signal vector The weight vector of the adaptive filter is updated using the normalized least mean square algorithm, and the update formula is as follows: in, For the updated weight vector, The step size factor is used to control the convergence speed and stability of the algorithm. To prevent regularization constants with denominators of zero.

[0009] Optionally, the instruction parsing engine includes a deep learning-based instruction recognition model. The instruction recognition model takes the acoustic features of the audio data after echo cancellation processing as input and performs joint decoding in conjunction with a KTV song selection domain biased language model. The KTV song selection domain biased language model is trained using a dedicated text corpus for the KTV song selection domain. The dedicated text corpus includes song names, singer names, song selection control instructions, and a set of common colloquial expressions of the instructions.

[0010] Optionally, the joint decoding process includes: The acoustic model of the instruction recognition model is obtained based on the input acoustic features. Calculated word sequence probability ; The domain-biased language model for KTV song selection is obtained as a word sequence. Assigned prior probabilities = ; By weighted fusion of the word sequence probabilities With the prior probability of the domain The logarithm of the search yields the word sequence that results in the highest fusion score. As a result of the identification, the decision logic is as follows: in, The fusion weights for the domain-biased language model. This indicates a search through all possible word sequences. The output is the sequence of words that maximizes the objective function. .

[0011] Optionally, the instruction recognition model includes: The acoustic feature extraction unit is configured to convert the audio data after echo cancellation processing into a frame-level Mel frequency cepstral coefficient feature sequence. A temporal feature encoder, connected to the acoustic feature extraction unit, includes multiple dilated causal convolutional layers, configured to perform deep temporal modeling on the Mel frequency cepstral coefficient feature sequence and output a high-dimensional context feature sequence. An attention decoder, connected to the temporal feature encoder, is configured to apply an attention mechanism to the high-dimensional context feature sequence, generate a sequence containing candidate lexical units, and calculate the probability of the word sequence. .

[0012] Optionally, the microphone receiver further includes a conflict arbitration module; the conflict arbitration module is configured to encapsulate the command signal stream into a data frame and forward it to the karaoke set-top box according to a preset arbitration strategy when it simultaneously receives command signal streams from multiple wireless microphones.

[0013] Optionally, the arbitration strategy includes: comparing the received signal strength indication values ​​corresponding to each instruction signal stream, and forwarding each instruction signal stream in descending order of the received signal strength indication values.

[0014] Optionally, the dedicated data transmission line adopts the USB or I2S communication protocol, and the karaoke set-top box sends a master clock signal to the microphone receiver through the dedicated data transmission line so that the audio data sent by the microphone receiver is synchronized with the sampling clock of the synchronous sampling module.

[0015] Secondly, this application discloses a KTV song selection method based on microphone voice command parsing, applicable to the KTV song selection system based on microphone voice command parsing as described in any of the preceding claims. The method includes the following steps: When the voice command button on the wireless microphone is triggered, the signal distribution circuit is activated, and the human voice signal collected by the wireless microphone is distributed in real time into a parallel amplification signal stream and a command signal stream. The wireless microphone transmits the command signal stream to the microphone receiver via a radio frequency link; the microphone receiver outputs the amplified signal stream to the audio amplification device through an audio output interface, and simultaneously encapsulates the received command signal stream into a data frame with channel status identifier and audio data through a data communication interface; The microphone receiver sends the data frame to the karaoke set-top box via a dedicated data transmission line; The karaoke set-top box determines the instruction parsing window based on the channel status identifier parsed from the data frame, and within the parsing window, it synchronously samples the background music digital signal output by the karaoke set-top box itself to generate an echo reference signal; The karaoke set-top box performs adaptive echo cancellation processing on the audio data parsed from the data frame based on the echo reference signal, performs semantic recognition on the processed audio data, and executes the corresponding song selection operation according to the recognition result.

[0016] Beneficial Effects: This application's KTV song selection system and method based on microphone voice command parsing separates human voice into independent amplified and command streams in real time at the signal source when the user triggers the voice command button. This eliminates false wake-up and misrecognition problems caused by environmental chat or other microphone singing sounds, ensuring control accuracy. Combined with the karaoke set-top box's synchronous sampling of background music and targeted adaptive echo cancellation processing, a dual anti-interference mechanism of front-end physical isolation and back-end precise noise reduction is constructed. This prevents background music and human voice mixing from contaminating the command signal at the source and effectively eliminates echoes coupled through air and circuits. This improves the clarity and signal-to-noise ratio of the voice command signal in the complex acoustic environment of a KTV, providing high-quality input for the language recognition of song selection commands and thus improving command recognition accuracy. Furthermore, this application eliminates the cumbersome step of repeatedly calling out voice wake-up words. Users only need to press a dedicated button on the microphone to directly issue commands, achieving natural and quick interaction, optimizing the user experience, and seamlessly integrating voice control into the KTV entertainment process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a structural block diagram of a KTV song selection system based on microphone voice command parsing provided in an embodiment of this application. Figure 2 This is a flowchart illustrating a KTV song selection method based on microphone voice command parsing, as provided in an embodiment of this application. Detailed Implementation

[0019] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0020] With the maturity of voice control technology, intelligent voice assistants have been attempted to be integrated into KTV song selection systems for convenient operation. However, KTV rooms, as unique acoustic environments, present a significant challenge to voice recognition technology due to the strong noise field created by the continuous accompaniment music, multiple singers singing, and conversations. Traditional voice interaction solutions designed for quiet home environments, when directly applied in these settings, reveal the following interconnected systemic problems: First, the command signal suffers irreversible contamination at the physical level. Existing solutions typically use the microphones in a karaoke room or ambient microphones to pick up the speech containing the command. This causes the user's song selection command to be linearly superimposed and nonlinearly coupled with the loud background music and other people's singing in the air and within the same audio hardware circuitry. This deep mixing that occurs at the signal source makes it difficult for any subsequent software noise reduction algorithm to completely separate it, resulting in an excessively low input signal-to-noise ratio for the speech recognition engine and making it difficult to guarantee recognition accuracy.

[0021] Secondly, the interactive logic of continuous listening is fundamentally contradictory to noisy, multi-person scenarios. The "hot word wake-up + continuous listening" mode, exemplified by smart speakers, requires the device to be constantly in a pickup state to capture the wake word. In a karaoke room, this can cause the system to be unable to distinguish between genuine song selection commands and song lyrics, casual conversations between friends, or even crosstalk from other rooms, resulting in frequent false wake-ups and misidentifications. This not only fails to provide reliable service but also interferes with the normal singing process, leading to a counterproductive user experience.

[0022] Finally, the cumbersome wake-up process disrupts the continuity of entertainment. In KTV scenarios where frequent song switching, selection, and parameter adjustments are required, each operation necessitates clearly calling out the device's wake-up word, such as "Xiao Ai Tongxue," which clashes with the relaxed and immersive atmosphere of entertainment. The long and unnatural interaction path leads to low usage of the voice function.

[0023] In summary, the core problem with existing technologies lies in the fact that noise in a KTV environment completely overlaps with the target command in time and space, and the interaction logic has not been reconstructed for scenarios involving multiple users, high noise levels, and high-frequency operations. Simply relying on backend signal processing algorithm optimization has reached its limit. This embodiment aims to provide a system-level solution that collaboratively innovates from the signal acquisition source to the interaction logic, enabling accurate, private, and smooth voice-based song selection interaction in complex acoustic environments.

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application. Secondly, in this document, the term "comprising" is 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 a process, method, article, or apparatus.

[0025] Firstly, this embodiment provides a KTV song selection system based on microphone voice command parsing for song selection, such as... Figure 1 As shown, the core of this system lies in its ability to automatically separate the user's voice into two independent signal streams at the signal source when the user issues a voice command by triggering a dedicated button on the wireless microphone. One signal stream is used for normal amplified singing, while the other serves as a clean command signal, which is sent directly to the karaoke set-top box for recognition via a newly added dedicated data transmission line. This design physically isolates the biggest sources of interference in the KTV environment (background music and other people's singing), thus solving the problems of false wake-up, high misrecognition rate, and the need for frequent wake-up words in traditional voice assistants in this scenario.

[0026] In detail, this system includes a wireless microphone, a microphone receiver, and a karaoke set-top box.

[0027] The wireless microphone includes a voice command button and a signal distribution circuit. The voice command button is a physical button located on the side of the microphone handle, which can be triggered by pressing or touch sensing. The user presses this button to select a song. The signal distribution circuit is an electronic circuit module inside the wireless microphone, configured to distribute the acquired human voice signal into parallel amplified signal streams and command signal streams in real time while the voice command button is triggered.

[0028] In practice, the pickup head of the wireless microphone (such as an electret condenser microphone) converts sound into an analog electrical signal. This signal, after passing through a preamplifier, is simultaneously input to two independent outputs of the signal distribution circuit. One output generates amplified signal streams, which are processed and transmitted according to the conventional wireless microphone workflow (such as processing and transmission for singing). The other output generates command signal streams, which are copies specifically prepared for subsequent command recognition. The signal distribution circuit can consist of an analog signal replication circuit (such as a voltage follower composed of operational amplifiers) and an array of electronic switches controlled by voice command buttons, ensuring that the two signals are generated simultaneously and independently when the button is triggered.

[0029] The microphone receiver is connected to the wireless microphone via an RF link and is configured to encapsulate the received command signal stream into a data frame with channel status identifier and audio data via a data communication interface, and to connect to an audio amplifier via an audio output interface to output an amplified signal stream.

[0030] In practice, the microphone receiver is typically a standalone hardware device, comprising an RF receiving module, a main control processor, and multiple physical interfaces. The RF link can employ wireless communication methods such as a proprietary 2.4GHz protocol or the UHF band. Once the microphone receiver receives the RF signal from the wireless microphone through its RF receiving module, the main control processor decodes it. The decoded amplified audio signal stream is directly output to an external amplifier or mixing console via a standard analog audio output interface (such as a 6.35mm TRS or XLR connector). For the decoded command signal stream, the main control processor encapsulates it, along with a channel status identifier (e.g., an 8-bit status byte) indicating that the microphone is currently in command input mode, into a complete data frame according to a specific communication protocol (e.g., adding a frame header before the audio data). This data frame is then prepared for transmission via a data communication interface (e.g., a USB Type-B interface or an RJ45 Ethernet port).

[0031] The karaoke set-top box connects to the microphone receiver's data communication interface via a dedicated data transmission cable. It includes a command parsing engine and a synchronous sampling module. The dedicated data transmission cable can be a USB cable or an audio cable conforming to the I2S protocol, used to establish an independent digital communication channel between the microphone receiver and the karaoke set-top box, distinct from traditional audio lines. The synchronous sampling module is a software functional module or hardware logic unit within the karaoke set-top box. It is configured to synchronously sample the background music digital signal output by the karaoke set-top box itself within the command parsing window, determined based on the channel status identifier parsed from the data frame, using it as an echo reference signal. The command parsing engine, also a software module within the karaoke set-top box, is configured to perform adaptive echo cancellation processing on the audio data parsed from the data frame based on the echo reference signal, and to perform semantic recognition on the processed audio data to execute the corresponding song selection operation.

[0032] In practice, when the parsed channel status identifier indicates that the current data frame carries a voice command, the karaoke set-top box immediately opens a brief command parsing window (e.g., lasting 3 seconds). Within this window, the synchronous sampling module captures the raw digital background music signal (PCM stream) being sent to the power amplifier from the karaoke set-top box's audio output bus (e.g., the ALSA audio driver layer) in real time and synchronously. The command parsing engine first calls the echo cancellation algorithm, using the synchronously sampled background music signal as a known reference input, to estimate and subtract the background music component from the received audio data (which is mixed with human voice commands and background music echoes). Then, the relatively pure human voice data is sent to the built-in speech recognition software for text conversion and semantic understanding, ultimately transforming it into executable song selection control commands such as "Play 'XX song'", "Choose song", and "Original version".

[0033] Based on the above, physical separation of the singing sound (amplified signal stream) and control commands (command signal stream) is achieved at the wireless microphone end through hardware button triggering. The command signal stream goes directly to the karaoke set-top box through a dedicated digital channel, avoiding mixing with the amplified signal in the analog audio link. At the same time, the karaoke set-top box uses precise synchronous sampling of its own output background music to provide a perfect reference signal for echo cancellation. The combination of these two aspects ensures the purity of the command signal input to the speech recognition engine to the greatest extent possible from both the signal source and signal processing stages. This enables the system to reliably extract user voice commands from extremely strong environmental noise and music interference, laying a solid foundation for subsequent high-accuracy recognition. Furthermore, users do not need to shout a wake word; they can directly issue commands by pressing a button, making the interaction more natural and convenient.

[0034] When a user presses the voice command button, their voice is simultaneously amplified and played back by the speakers. This can lead to a contradiction: the user's private control commands (such as "change song") are amplified, interfering with other users singing, or others' chat content may accidentally trigger command recognition. Therefore, in this embodiment, the wireless microphone also includes a mute control module controlled by the voice command button; the mute control module is configured to attenuate or block the amplified signal stream when the voice command button is pressed, so that the human voice signal is transmitted to the karaoke set-top box only through the command signal stream.

[0035] In practical implementation, the mute control module can be integrated into the circuit path after the signal distribution circuit and before the RF transmission module of the wireless microphone. It can be an analog switch (such as a MOSFET switch) directly controlled by a voice command button, or a digitally controlled attenuator. When the voice command button is not pressed, this module has no effect on the amplified signal flow path. Once the voice command button is pressed, the module immediately activates: in blocking mode, the analog switch disconnects, completely cutting off the amplified signal flow path; in attenuation mode, the digitally controlled attenuator significantly reduces the gain of the amplified signal flow (e.g., by 60dB), achieving a near-silent effect. Thus, the user's voice during button presses will not be sent to the audio equipment for amplification, but will instead be silently transmitted to the karaoke set-top box via the command signal flow.

[0036] Based on the above, the mute control module serves a dual purpose: ensuring the privacy of commands and preventing accidental triggering. It guarantees that user control commands will not be publicly played, avoiding disturbance to others, and also prevents other ambient sounds from being accidentally picked up and identified through the amplification channel. This makes voice command operation a personalized and quiet interaction method, further enhancing the system's usability and user experience.

[0037] To achieve high-performance echo cancellation and ensure the clarity of the final voice command, the adaptive echo cancellation process in this embodiment includes: Obtain the echo reference signal vector corresponding to the echo reference signal. ; Echo reference signal vector Input to a weighted vector An adaptive filter, where the weight vector Used for the echo reference signal vector Weighting is applied to simulate the echo path and generate an echo estimation signal; The residual signal is obtained by subtracting the echo estimation signal from the audio data of the data frame. , and serve as the target speech signal after echo cancellation; Based on residual signal With echo reference signal vector The normalized least mean square algorithm is used to update the weight vector of the adaptive filter. The update formula is as follows: in, For the updated weight vector, The step size factor is used to control the convergence speed and stability of the algorithm. To prevent regularization constants with denominators of zero.

[0038] In practical implementation, it is assumed that at discrete time points The background music signal sample value obtained by the synchronous sampling module is Echo reference signal vector It consists of the present and the past A column vector consisting of the reference signal sample values ​​at each time step, i.e. superscript This represents the transpose of a vector. The length of the filter is set according to the duration of the room echo, for example, corresponding to a 200ms echo at a 48kHz sampling rate. A value of 9600 is acceptable. (Weight vector) These are the coefficients of the adaptive filter, typically initialized as a zero vector. Echo estimation signal. Through calculation and The inner product is obtained as follows: Among them, the summation index This represents the index of an element in the weight vector and reference signal vector, with a value range of [0,1]. It is derived from the received audio data containing echoes. Subtracting this estimated value from the audio sample values ​​in the data frame yields the residual signal: This residual signal is the desired approximation of the target speech signal after eliminating background music echoes, and it also serves as error feedback. The normalized least mean square algorithm uses this error to update the weight vector. In the update formula of the normalized least mean square algorithm, Representing vectors The square of the Euclidean norm, i.e. Sum of squares of each component. Step factor. A decimal number between 0 and 2 (e.g., 0.1) is typically chosen to control the magnitude of the update. Regularization constant. It is a very small positive number (e.g., = ), used in When the value is very small, avoid the denominator being zero, which could lead to unstable updates.

[0039] Based on the above, by continuously adjusting the weight vector, the acoustic echo path from the speaker to the microphone receiver is adaptively simulated, making the echo estimation signal output by the filter approximate the real echo as closely as possible, thereby maximally canceling the echo component in the residual signal. Furthermore, the algorithm used has low computational complexity, fast convergence speed, and good stability to changes in the energy of the input signal, effectively extracting pure speech commands from the mixed signal.

[0040] The audio data obtained after echo cancellation processing needs to be converted into specific song selection instructions. However, general speech recognition models have a low recognition rate for KTV-specific vocabulary such as song titles and artist names. Therefore, in this embodiment, the instruction parsing engine includes a deep learning-based instruction recognition model. The instruction recognition model takes the acoustic features of the echo-cancelled audio data as input and performs joint decoding in conjunction with a KTV song selection domain biased language model. The KTV song selection domain biased language model is trained using a dedicated text corpus for the KTV song selection domain. This dedicated text corpus includes song titles, artist names, song selection control instructions, and a set of common colloquial expressions for the instructions.

[0041] In practical implementation, the instruction recognition model can adopt an end-to-end deep learning architecture, such as a Transformer-based model or an RNN-T model. Its acoustic feature input is typically Mel-frequency cepstral coefficients or Filterbank features. The domain-biased language model is an N-gram language model or neural language model specifically trained for KTV scenarios. The dedicated text corpus for training this model needs to be collected on a large scale, and its sources can include: ① a structured list of songs and singers exported from a music database; ② historical song request logs from KTV karaoke systems (after anonymization); ③ manually simulated, colloquial song request texts, such as "Please request Jay Chou's 'Sunny Day'", "Please play 'Confession Balloon'", "Cut the song", "Replay", "Please prioritize this song", etc. By training on these corpora, the language model learns the co-occurrence probabilities between words in KTV scenarios. For example, "Jay Chou" is likely followed by song titles such as "Sunny Day" and "Seven Mile Fragrance", while "cut" is likely followed by "song". This injection of domain knowledge provides strong prior constraints for the decoding process.

[0042] Based on the above, during joint decoding, the acoustic model is responsible for calculating the probability of which phonemes / words the heard audio might correspond to, while the domain-biased language model is responsible for assessing the likelihood that "these words, according to the habits of a KTV setting, can be linked together to form a reasonable instruction." The combination of these two models can correct errors that the acoustic model might produce due to noise or accents, improving the accuracy of recognizing proper nouns and colloquial instructions. This allows the system to more reliably understand complex user instructions such as "select a song with rare characters" or "play a song by Jolin Tsai."

[0043] To achieve effective utilization of domain knowledge, in this embodiment, the joint decoding process includes: The acoustic model for obtaining the instruction recognition model is based on the input acoustic features. Calculated word sequence probability ; Obtain the domain bias language model for KTV song selection as a word sequence Assigned prior probabilities = ; By weighted fusion word sequence probability Prior probability of the domain The logarithm of the search yields the word sequence that results in the highest fusion score. As a result of the identification, the decision logic is as follows: in, For the fusion weights of the domain-biased language model, This indicates a search through all possible word sequences. The output is the sequence of words that maximizes the objective function. .

[0044] In practical implementation, input acoustic features This represents a sequence of preprocessed and feature-extracted speech features. The acoustic model (usually a deep learning model) will process all possible word sequences. Calculate a conditional probability , indicating that when the feature is observed Under the premise that the word sequence is The possibility of this. Meanwhile, the domain-biased language model will, based on its training data, generate a sequence of candidate words for each word. Calculate an independent prior probability This indicates the prevalence of the word sequence in a KTV song selection scenario. (Fusing weight) It is a hyperparameter used to balance the confidence level between acoustic evidence and language priors, and its value can be determined by tuning on a development set (e.g., Set to a value between 0.5 and 1.5). Decoding the search process (i.e. Typically, a beam search algorithm is used. This algorithm, when expanding the candidate sequence at each step, considers not only the acoustic score but also the prior score obtained by querying the language model based on the currently generated word sequence, multiplied by [the previous score]. Ultimately, the complete word sequence with the highest overall score is selected as the recognition result. .

[0045] Based on the above, the acoustic model and language model are jointly optimized within a unified probabilistic framework. This allows the system to rely more on domain common sense to make reasonable judgments when the acoustic signal is ambiguous, while avoiding excessive constraints from the language model when the acoustic signal is very clear. This effectively improves the system's overall robustness and recognition accuracy to complex, proprietary instructions.

[0046] To achieve the mapping from audio features to word sequence probabilities and effectively capture long-term dependencies in speech, in this embodiment, the instruction recognition model includes: The acoustic feature extraction unit is configured to convert the audio data after echo cancellation processing into a frame-level Mel frequency cepstral coefficient feature sequence. The temporal feature encoder, connected to the acoustic feature extraction unit, includes multiple dilated causal convolutional layers configured to perform deep temporal modeling on the Mel frequency cepstral coefficient feature sequence and output a high-dimensional context feature sequence. The attention decoder, connected to the temporal feature encoder, is configured to apply an attention mechanism to the high-dimensional context feature sequence, generating a sequence containing candidate words and calculating the probability of the word sequence. .

[0047] In practical implementation, the acoustic feature extraction unit first performs pre-emphasis, framing (frame length 25ms, frame shift 10ms), and windowing on the echo-cancelled audio sampled at 16kHz. Then, it calculates 40-dimensional Mel-frequency cepstral coefficient features for each frame, ultimately obtaining a feature derived from... Sequence of frame features ,in The total number of frames is given. A temporal feature encoder can be constructed by stacking multiple dilated causal convolutional blocks. Each block contains a one-dimensional causal convolutional layer, a dilation factor, weight normalization, gated activation units, and residual connections. For example, eight such blocks can be used, with the dilation factor doubling layer by layer (1,=2,=4,=...,=128). This design allows for a very large receptive field at the top layer (e.g., exceeding 1 second), enabling the capture of long-range contextual information in the speech signal while maintaining structural causality (i.e., the output is independent of future inputs). The encoder ultimately processes the input feature sequence... Mapped to a high-dimensional context feature sequence An attention-based decoder is typically an autoregressive Transformer decoder or an RNN decoder that generates an output sequence word-by-word (e.g., Chinese characters or sub-word units). When generating each new word, the decoder dynamically selects from the contextual feature sequence output by the encoder through an attention mechanism. The algorithm focuses on the most relevant parts and combines them with previously generated historical lexical information to predict the probability distribution of the next lexical term. By connecting the lexical probabilities generated by the decoder at each step, the entire word sequence can be obtained. The approximate probability, where Given the length of the word sequence, the approximate probability can be calculated as follows: = To simplify the expression, let Indicates position All previously generated word sequences Then the approximate probability can be calculated as follows: =， Each of them This means that the attention decoder is based on the generated historical sequence. and all acoustic context features Under the condition of predicting the next word element, The probability of.

[0048] Based on the above, the instruction recognition model efficiently models the long-term temporal pattern of speech signals through dilated causal convolution, and then achieves flexible alignment between encoded information and the decoding process through an attention mechanism. The combination of these two aspects enables the acoustic model to more accurately capture phoneme and word boundary information in continuous speech streams, providing high-quality acoustic likelihood scores for subsequent joint decoding, which is the foundation for the high performance of the entire speech recognition system.

[0049] In multi-user KTV scenarios, multiple users may simultaneously press the voice command buttons on their respective microphones. If multiple command signal streams simultaneously flood the karaoke set-top box, it can lead to data conflicts, system inability to process, or erratic command execution. Therefore, in this embodiment, the microphone receiver also includes a conflict arbitration module; the conflict arbitration module is configured to, when simultaneously receiving command signal streams from multiple wireless microphones, encapsulate the command signal streams into data frames according to a preset arbitration strategy and forward them to the karaoke set-top box.

[0050] In its implementation, the conflict arbitration module is a software logic module running within the microphone receiver's main control processor. When the RF receiving module decodes command signal streams from two or more wireless microphones within a very short time window (e.g., 10 milliseconds), the main control processor triggers the conflict arbitration logic. This module's responsibility is to sort or select these concurrent command signal streams according to a predetermined strategy, then encapsulate them sequentially into data frames and send them to the karaoke set-top box via a dedicated data transmission line. Furthermore, this forwarding refers to ordered, non-simultaneous transmission to avoid data collisions on the bus.

[0051] Based on the above, the role of the conflict arbitration module is to manage concurrent requests from multiple users, resolve and schedule potential physical layer or data link layer conflicts at the microphone receiver as an intermediate node, thereby ensuring that instruction data can be delivered to the karaoke set-top box for processing in an orderly and reliable manner, and maintaining the stability and controllability of the system.

[0052] To ensure that the arbitration strategy is fair and easy to implement, in this embodiment, the arbitration strategy includes: comparing the received signal strength indication values ​​corresponding to each instruction signal stream, and forwarding each instruction signal stream in descending order of the received signal strength indication values.

[0053] In practical implementation, the signal strength of a wireless microphone is stable within a certain range when transmitting radio frequency signals. The RF receiving module of the microphone receiver can measure the Received Signal Strength Indication (RSSI) in real time when demodulating each wireless microphone signal. RSSI is a relative value reflecting signal strength; generally, the closer the signal source is to the receiver and the fewer obstacles, the higher the RSSI value (the smaller the negative value). When a collision occurs, the collision arbitration module reads the instantaneous RSSI value corresponding to each concurrent command signal stream. Then, these command signal streams are placed in a priority queue, with the signal stream having the highest RSSI value assigned the highest priority, and its data is encapsulated and transmitted first. After the highest priority data frame is transmitted, the next highest priority frame is transmitted, and so on. If the RSSI values ​​of two signal streams are very close, other discrimination conditions can be added, such as the microphone's physical ID number.

[0054] Based on the above, an arbitration strategy based on the received signal strength indicator (RSI) is adopted. The principle is to prioritize microphone commands from users closer to the receiver, whose signal quality is likely better. This is because users closer to the receiver may have stronger intentions, and their voice commands are less susceptible to wireless interference during transmission, thus ensuring the system responds first to the clearest command. This strategy is simple to implement, requires no additional communication overhead, and effectively resolves concurrency issues.

[0055] The microphone receiver and the karaoke set-top box are two independent devices, and their internal clock sources may have slight differences (clock drift). If their sampling is out of sync, the karaoke set-top box will experience a significant decrease in cancellation performance when processing audio data (especially when cancelling the synchronously sampled echo reference signal) due to time axis misalignment. Therefore, in this embodiment, the dedicated data transmission line uses the USB or I2S communication protocol, and the karaoke set-top box sends a master clock signal to the microphone receiver through the dedicated data transmission line to synchronize the audio data sent by the microphone receiver with the sampling clock of the synchronous sampling module.

[0056] In practical implementation, if the I2S protocol is used, it includes a separate serial clock line (SCK) and a word clock line (LRCK). The karaoke set-top box, acting as the I2S master device, sends a precise master clock signal to the microphone receiver, acting as the slave device, via these two lines. The microphone receiver uses this master clock to drive the sampling rhythm of its internal analog-to-digital converter and organizes the transmission of audio data according to this clock. If a USB audio protocol is used, the karaoke set-top box (USB master) synchronizes the clock of the slave device (microphone receiver) using the timestamp information in the USB data packet. The microphone receiver needs to adjust its sampling rate to match the master clock based on the synchronization information sent by the master. In this way, regardless of the accuracy of the microphone receiver's local crystal oscillator, the sampling time of the audio data it generates is strictly aligned with the system master clock of the karaoke set-top box.

[0057] Based on the above, the role of clock synchronization is to ensure that the microphone receiver and the karaoke set-top box are consistent in terms of time base. This allows the karaoke set-top box's synchronization sampling module to accurately know the temporal correspondence between each audio data sampling point sent by the microphone receiver and its own generated background music signal. Only under this premise can the adaptive echo cancellation algorithm achieve efficient echo cancellation and ensure the input quality of subsequent speech recognition.

[0058] Secondly, this embodiment provides a KTV song selection method based on microphone voice command parsing, which is applied to the KTV song selection system based on microphone voice command parsing as described above, such as... Figure 2 As shown, the method includes the following steps in sequence: When the voice command button on the wireless microphone is triggered, the signal distribution circuit is activated, which distributes the human voice signal collected by the wireless microphone into a parallel amplified signal stream and a command signal stream in real time. The wireless microphone sends the command signal stream to the microphone receiver via the radio frequency link; the microphone receiver outputs the amplified signal stream to the audio amplification device through the audio output interface, and at the same time encapsulates the received command signal stream into a data frame with channel status identifier and audio data through the data communication interface; The microphone receiver transmits data frames to the karaoke set-top box via a dedicated data transmission line; The karaoke set-top box determines the instruction parsing window based on the channel status identifier parsed from the data frame, and within the parsing window, it synchronously samples the background music digital signal output by the karaoke set-top box itself to generate an echo reference signal; The karaoke set-top box uses an echo reference signal to perform adaptive echo cancellation processing on the audio data parsed from the data frame, and performs semantic recognition on the processed audio data. Based on the recognition results, it executes the corresponding song selection operation.

[0059] It should be noted that the KTV song selection method based on microphone voice command parsing in this embodiment corresponds to the aforementioned KTV song selection system based on microphone voice command parsing. Therefore, the parts of this KTV song selection method based on microphone voice command parsing that are not described in detail (including but not limited to specific technical means and effects) can be referred to the relevant descriptions in the aforementioned KTV song selection system based on microphone voice command parsing, and will not be repeated here.

[0060] In the embodiments provided in this application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the associated hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available medium accessible to a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0061] Furthermore, the technical solution involved in this embodiment will collect and process user voice data for command recognition during implementation. It is hereby solemnly declared that all voice data collection and processing shall strictly comply with relevant personal information protection laws, regulations, and standards. In specific product implementations, the system should clearly and explicitly inform the user that voice data will be collected and used for song selection command recognition upon first use or initial activation of the relevant function, and obtain the user's explicit authorization. Voice data processing should primarily be completed on local devices (such as song selection set-top boxes), and only when necessary (e.g., to update the language model) with the user's additional consent may anonymize and desensitize the data before uploading it to the cloud. The system should provide clear settings options, allowing users to disable the voice command function or manage their voice data at any time. The technical solution described in this specification assumes by default that all data processing activities are conducted on a legal basis with the user's knowledge and consent.

[0062] Finally, it should be noted that the above description is only a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A KTV song selection system based on microphone voice command parsing for song selection, characterized in that, include: A wireless microphone includes a voice command button and a signal distribution circuit; the signal distribution circuit is configured to distribute the acquired human voice signal into a parallel amplified signal stream and a command signal stream in real time during the period when the voice command button is triggered. A microphone receiver is connected to the wireless microphone via a radio frequency link and is configured to encapsulate the received command signal stream into a data frame with channel status identifier and audio data via a data communication interface, and to connect to an audio amplifier via an audio output interface to output the amplified signal stream. The karaoke set-top box is connected to the data communication interface of the microphone receiver via a dedicated data transmission line. It includes an instruction parsing engine and a synchronous sampling module. The synchronous sampling module is configured to synchronously sample the background music digital signal output by the karaoke set-top box itself within an instruction parsing window determined according to the channel status identifier parsed from the data frame, so as to use it as an echo reference signal. The instruction parsing engine is configured to perform adaptive echo cancellation processing on the audio data parsed from the data frame based on the echo reference signal, and to perform semantic recognition on the processed audio data to execute the corresponding song selection operation.

2. The KTV song selection system based on microphone voice command parsing according to claim 1, characterized in that, The wireless microphone also includes a mute control module controlled by the voice command button; the mute control module is configured to attenuate or block the amplified signal stream when the voice command button is pressed, so that the human voice signal is transmitted to the karaoke set-top box only through the command signal stream.

3. The KTV song selection system based on microphone voice command parsing according to claim 1, characterized in that, The adaptive echo cancellation process includes: Obtain the echo reference signal vector corresponding to the echo reference signal. ; The echo reference signal vector Input to a weighted vector An adaptive filter, wherein the weight vector Used for the echo reference signal vector Weighting is applied to simulate the echo path and generate an echo estimation signal; The residual signal is obtained by subtracting the echo estimation signal from the audio data of the data frame. , and serve as the target speech signal after echo cancellation; According to the residual signal With the echo reference signal vector The weight vector of the adaptive filter is updated using the normalized least mean square algorithm, and the update formula is as follows: in, For the updated weight vector, The step size factor is used to control the convergence speed and stability of the algorithm. To prevent regularization constants with denominators of zero.

4. The KTV song selection system based on microphone voice command parsing according to claim 1, characterized in that, The instruction parsing engine includes a deep learning-based instruction recognition model. The instruction recognition model takes the acoustic features of the audio data after echo cancellation processing as input and performs joint decoding in conjunction with a KTV song selection domain bias language model. The KTV song selection domain bias language model is trained using a dedicated text corpus for the KTV song selection domain. The dedicated text corpus includes song names, singer names, song selection control instructions, and a set of common colloquial expressions of the instructions.

5. The KTV song selection system based on microphone voice command parsing according to claim 4, characterized in that, The joint decoding process includes: The acoustic model of the instruction recognition model is obtained based on the input acoustic features. Calculated word sequence probability ; The domain-biased language model for KTV song selection is obtained as a word sequence. Assigned prior probabilities = ; By weighted fusion of the word sequence probabilities With the prior probability of the domain The logarithm of the search yields the word sequence that results in the highest fusion score. As a result of the identification, the decision logic is as follows: in, The fusion weights for the domain-biased language model. This indicates a search through all possible word sequences. The output is the sequence of words that maximizes the objective function. .

6. The KTV song selection system based on microphone voice command parsing according to claim 5, characterized in that, The instruction recognition model includes: The acoustic feature extraction unit is configured to convert the audio data after echo cancellation processing into a frame-level Mel frequency cepstral coefficient feature sequence. A temporal feature encoder, connected to the acoustic feature extraction unit, includes multiple dilated causal convolutional layers, configured to perform deep temporal modeling on the Mel frequency cepstral coefficient feature sequence and output a high-dimensional context feature sequence. An attention decoder, connected to the temporal feature encoder, is configured to apply an attention mechanism to the high-dimensional context feature sequence, generate a sequence containing candidate lexical units, and calculate the probability of the word sequence. .

7. The KTV song selection system based on microphone voice command parsing according to claim 1, characterized in that, The microphone receiver also includes a conflict arbitration module; the conflict arbitration module is configured to encapsulate the command signal stream into a data frame and forward it to the karaoke set-top box according to a preset arbitration strategy when it simultaneously receives command signal streams from multiple wireless microphones.

8. The KTV song selection system based on microphone voice command parsing according to claim 7, characterized in that, The arbitration strategy includes: comparing the received signal strength indication values ​​corresponding to each instruction signal stream, and forwarding each instruction signal stream in descending order of received signal strength indication values.

9. The KTV song selection system based on microphone voice command parsing according to claim 1, characterized in that, The dedicated data transmission line adopts the USB or I2S communication protocol, and the karaoke set-top box sends a master clock signal to the microphone receiver through the dedicated data transmission line so that the audio data sent by the microphone receiver is synchronized with the sampling clock of the synchronous sampling module.

10. A KTV song selection method based on microphone voice command parsing, applied to the KTV song selection system based on microphone voice command parsing as described in any one of claims 1-9, characterized in that, The method includes the following steps: When the voice command button on the wireless microphone is triggered, the signal distribution circuit is activated, and the human voice signal collected by the wireless microphone is distributed in real time into a parallel amplification signal stream and a command signal stream. The wireless microphone transmits the command signal stream to the microphone receiver via a radio frequency link; the microphone receiver outputs the amplified signal stream to the audio amplification device through an audio output interface, and simultaneously encapsulates the received command signal stream into a data frame with channel status identifier and audio data through a data communication interface; The microphone receiver sends the data frame to the karaoke set-top box via a dedicated data transmission line; The karaoke set-top box determines the instruction parsing window based on the channel status identifier parsed from the data frame, and within the parsing window, it synchronously samples the background music digital signal output by the karaoke set-top box itself to generate an echo reference signal; The karaoke set-top box performs adaptive echo cancellation processing on the audio data parsed from the data frame based on the echo reference signal, performs semantic recognition on the processed audio data, and executes the corresponding song selection operation according to the recognition result.

Images