Intelligent microphone starting method based on voice recognition

Abstract

The invention discloses an intelligent microphone starting method based on voice recognition, and relates to the technical field of voice recognition. The method comprises the following steps: acquiring a user voice signal and surrounding environment audio data to obtain a mixed audio signal stream; identifying an abnormal waveform and determining a frequency feature vector; determining a waveform fluctuation amplitude and obtaining an enhanced waveform stable representation; recognizing syllable interval duration and analyzing interval shortening degree; analyzing the urgent speech speed characteristics to determine a time domain stretching compensation value, and adjusting the syllable interval duration to obtain a syllable sequence; identifying spectrum features of the environment background audio and determining candidate wake-up words; performing similarity matching with an emergency wake-up word template to determine an intention recognition result; and analyzing the confidence score to obtain an emergency response instruction execution signal. Through multi-level signal processing and intelligent feature extraction, high-precision voice intention recognition in a complex environment is realized, the robustness and reliability of an emergency response system are remarkably improved, and rapid help seeking of a user in a crisis scene is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of speech recognition technology, and in particular relates to a method for activating a smart microphone based on speech recognition. Background Technology

[0002] In the field of intelligent voice interaction, voice wake-up technology, as the core entry point for human-computer interaction, bears the important mission of efficient communication between users and devices. Especially in emergency situations, the reliability and accuracy of the voice wake-up system directly affect user safety and rescue efficiency. However, when users are under stress or in abnormal environments, physiological changes caused by fear, tension, and other emotions can significantly affect voice characteristics. For example, a user's voice may tremble, their speech rate may increase, or their pronunciation may become distorted; these non-standard inputs often prevent the system from accurately recognizing the voice. Furthermore, environmental factors such as sudden noise or drastic changes in background sound can also interfere with voice signal capture, making it difficult for the system to distinguish between valid wake-up commands and environmental interference.

[0003] At the technical level, the core challenge of voice wake-up systems lies in handling the dynamic changes in speech signals under stress. A user's voice tremor causes abnormal fluctuations in the frequency waveform. These fluctuations not only affect the clarity of syllables but also alter the phoneme structure of the wake-up word, making it difficult for the system to match the preset speech template. Simultaneously, rapid speech shortens the syllable intervals, leading to significant changes in the temporal characteristics of the speech signal, further increasing the recognition difficulty. These dynamically changing speech features intertwine with abnormal sounds in the environment, such as sudden sharp noises or rapid switching of background sounds, requiring the system to accurately extract effective user commands from multimodal information. Abnormal fluctuations in the frequency waveform of voice tremor are the primary technical challenge.

[0004] Because users' vocal cord control decreases under stress, the frequency distribution of speech signals exhibits irregular changes, making it difficult for traditional static template matching methods to adapt to this dynamic nature. For example, in a fire or emergency, a user might utter a trembling cry for help due to tension, but the system may fail to respond promptly because it cannot recognize this non-standard pronunciation. This dynamic waveform change directly leads to another technical challenge: the rapid variation in syllable intervals. Rapid speech speed drastically shortens the time interval between syllables, requiring the system to complete phoneme segmentation and semantic parsing within a very short time; otherwise, it may miss crucial instructions.

[0005] Therefore, accurately capturing and analyzing the dynamic characteristics of tremor frequency waveforms and rapid syllable intervals in a user's voice under stress, while filtering out interference from abnormal sounds in complex environments, has become a key issue for intelligent voice wake-up systems to improve response reliability in emergency help scenarios. The core of this problem lies in the system's ability to quickly and accurately determine the user's intent to seek help even under conditions of significant physiological changes in the user's voice and severe environmental interference. For example, when a user issues a trembling and rapid wake-up command in a noisy public place due to a sudden event, the system must comprehensively analyze the voice waveform, syllable duration, and environmental noise characteristics within a very short time to ensure that the response opportunity is not missed due to signal abnormalities. This not only requires the system to have the ability to dynamically adapt to voice changes but also to maintain high robustness in the fusion processing of multimodal information, which is a key problem that this research urgently needs to solve. Summary of the Invention

[0006] The purpose of this invention is to provide a smart microphone activation method based on speech recognition. Through multi-level signal processing and intelligent feature extraction, it achieves high-precision speech intent recognition in complex environments, significantly improving the robustness and reliability of emergency response systems and ensuring users can quickly seek help in crisis scenarios. This invention can be achieved through the following technical solutions: This application provides a method for activating a smart microphone based on speech recognition, including: User voice signals and ambient audio data are collected, and a microphone array is used to capture multi-channel input signals to obtain a mixed audio signal stream. Frequency waveform features of the mixed audio signal stream are extracted, and abnormal waveforms are identified. Adaptive filtering is applied to the abnormal waveforms to determine the filtered frequency feature vector. Based on the filtered frequency feature vector, waveform fluctuation amplitude is determined. If the waveform fluctuation amplitude exceeds a preset fluctuation threshold, it is identified as a tremor-related frequency band and amplified to obtain an enhanced waveform stability representation. The enhanced waveform stability representation of the target time period is analyzed to identify syllable interval duration, and the degree of interval shortening is analyzed based on the syllable interval duration. The process is repeated. The rapid speech rate characteristics are analyzed to determine the temporal stretching compensation value. The syllable interval duration is adjusted based on this value to obtain a temporally adjusted syllable sequence. The surrounding environmental audio data is analyzed to identify the spectral characteristics of the background audio. Combined with the temporally adjusted syllable sequence, environmental interference is eliminated and residual noise is processed to determine a clean candidate wake-up word. The clean candidate wake-up word is then matched with a preset emergency wake-up word template to determine the intent recognition result. The confidence score of the intent recognition result is analyzed. If the confidence score exceeds the response trigger condition, an emergency response assistance process is triggered, resulting in an emergency response command execution signal.

[0007] The beneficial effects of this invention are as follows: Addressing the challenge of accurately recognizing rapid speech and trembling characteristics in complex environments, this invention uses a multi-channel microphone array to collect speech and environmental audio. After fusing the initial signal, it extracts frequency waveform features using a convolutional neural network and incorporates adaptive filtering to eliminate abnormal fluctuations. For cases where the fluctuation amplitude in the trembling frequency band exceeds a threshold, an attention mechanism is used to weight and enhance waveform stability. Simultaneously, it analyzes rapid speech rate characteristics using syllable intervals to generate a time-domain adjusted syllable sequence. Furthermore, this invention integrates an environmental noise model, eliminates residual interference through secondary convolution processing, forms a clean wake-up word candidate sequence, and utilizes an attention mechanism to integrate dynamic context matching of emergency command templates, ultimately triggering the help-seeking process. Through multi-level signal processing and intelligent feature extraction, this invention achieves high-precision speech intent recognition in complex environments, significantly improving the robustness and reliability of emergency response systems and ensuring users can quickly seek help in crisis scenarios. Attached Figure Description

[0008] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0009] Figure 1 A flowchart illustrating the steps of a smart microphone activation method based on speech recognition, provided in an embodiment of this application. Detailed Implementation

[0010] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0011] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to any and all possible combinations comprising one or more of the associated listed items.

[0012] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0013] Please see Figure 1 This application provides a method for activating a smart microphone based on speech recognition, comprising the following steps: Step S101: Real-time acquisition of user voice signals and ambient audio data; capture of multi-channel input signals using a microphone array to obtain a mixed audio signal stream containing timestamps.

[0014] Using ambient audio data, a beamforming algorithm is employed to process the multi-channel input signals of the microphone array. This algorithm uses the multi-channel input signals as input and calculates the phase and amplitude differences of each channel to enhance signals in specific directions, resulting in a directionally enhanced speech signal. For this directionally enhanced speech signal, a real-time timestamp is obtained, yielding an annotated mixed audio signal stream. Based on this annotated mixed audio signal stream, the ambient noise level is determined. This level is obtained by calculating the average energy value of the non-speech components in the annotated mixed audio signal stream. If the noise level exceeds a preset threshold, noise suppression processing is applied. This noise suppression process involves subtracting the noise spectrum from the annotated mixed audio signal stream using spectral subtraction, resulting in a purified audio stream. User speech features are extracted from the purified audio stream. These features are obtained by calculating the Mel-frequency cepstral coefficients in the purified audio stream, determining the final signal recognition result.

[0015] Specifically, in one implementation, audio signal capture is achieved by real-time acquisition of user voice signals and ambient audio data.

[0016] Specifically, the voice wake-up system is equipped with a microphone array, consisting of multiple microphone units arranged in a linear or circular structure to capture audio signals from different directions. This array can utilize the phase difference of sound wave propagation to enhance the signal. For example, when the user speaks, the microphone array simultaneously receives the signal and adjusts the gain of each channel using internal algorithms to emphasize the sound from the direction of the user's voice. Ambient audio data, including background noise such as indoor echoes or outdoor wind sounds, is mixed with the user's voice to form the initial multi-channel input signal.

[0017] It should be noted that the microphone array capture process ensures the spatiotemporal resolution of the signal, improving the accuracy of subsequent processing. Furthermore, the multi-channel input signals captured by the microphone array undergo preliminary processing to generate a mixed audio signal stream.

[0018] In one possible implementation, each microphone channel independently acquires audio samples at a high sampling rate to ensure real-time performance. During acquisition, the system monitors signal strength in real time and applies beamforming technology, which focuses on specific sound sources by calculating the time delay difference between the signals from each channel. For example, in a smart home scenario, when a user speaks in the living room, the array prioritizes amplifying speech from the direction of the sofa while suppressing background noise from the kitchen. This processing ensures that the mixed signal stream includes the main component of the user's speech and auxiliary information from the ambient audio, demonstrating the versatility of the technology. In another embodiment, for a conference room environment, the array can be configured in a ring layout to capture the voice signals of multiple users, ensuring that the signal stream covers the entire space.

[0019] Preferably, the generated mixed audio signal stream includes timestamps to achieve time synchronization of the signals.

[0020] Specifically, timestamps are generated using the system clock, such as precise time stamps synchronized using the NTP protocol, with a millisecond-level timestamp appended to each frame of audio data. This mechanism ensures the orderliness of the signal stream, facilitating subsequent analysis. For example, in voice assistant applications, timestamps help distinguish consecutive voice commands, avoiding confusion caused by signal overlap.

[0021] In one embodiment, the timestamp addition is performed immediately after signal acquisition, implemented through a hardware clock chip with an accuracy down to the microsecond level, thereby supporting high real-time requirements.

[0022] For example, in actual operation, the microphone array acquisition process can be divided into several stages. First, during array initialization, the position and sensitivity of each microphone are calibrated to ensure consistency between channels. Then, the real-time acquisition stage begins; after the user's voice signal is detected by the array, the system calculates the azimuth angle of the sound source and adjusts the signal weights using a phase difference formula.

[0023] For example, if the distance between two microphones is 10 centimeters, the difference in sound wave delay can be used to locate the sound source. This detailed capture mechanism not only improves signal quality but also provides robustness in audio processing. In one implementation, for noisy environments such as offices, the array can integrate adaptive filters to further separate user speech and ambient audio, generating a cleaner mixed signal stream.

[0024] Understandably, the versatility of this application is reflected in various scenarios. For example, in in-vehicle voice systems, the microphone array captures driver commands while simultaneously recording road noise, forming a timestamped signal stream for subsequent speech recognition. This implementation maintains domain uniformity and avoids cross-domain expansion. Furthermore, the generation process of the mixed audio signal stream emphasizes real-time performance, managing signal delay through buffers to ensure a continuous and uninterrupted output stream.

[0025] In one embodiment, the system can be extended to portable devices, such as microphone arrays on smartphones, to capture user voice and ambient audio, generating a signal stream containing timestamps. The process includes digitizing the array signal, using an ADC chip to convert the analog signal into a digital stream, and immediately appending a timestamp. This approach performs well in mobile scenarios and can handle dynamic environmental changes.

[0026] It should be noted that, through the above steps, this application achieves efficient audio data acquisition, providing a stable signal foundation for voice interaction applications and supporting more precise subsequent processing. Furthermore, in another possible implementation, the microphone array can be configured with different numbers of channels to adapt to environments of varying sizes. The acquired multi-channel signals are mixed using a weighted averaging method to form a single audio stream, and a timestamp sequence is embedded. This flexibility enhances the adaptability of the technology.

[0027] For example, the addition of timestamps is not only used for synchronization but also to assist in anomaly detection. For instance, if the timestamp interval in the signal stream is abnormal, the system can trigger a resampling mechanism to ensure data integrity. In one implementation, the output of the entire process is a storable mixed audio signal stream, suitable for long-term monitoring scenarios, such as audio recording in intelligent surveillance systems.

[0028] Step S102: Extract the frequency waveform features of the mixed audio signal stream, identify abnormal waveforms based on the frequency waveform features, perform adaptive filtering on the abnormal waveforms, and determine the filtered frequency feature vector.

[0029] Frequency waveform features are obtained from the mixed audio signal stream using Fourier transform, and the peak distribution in these features is determined by comparing the amplitude differences between adjacent frequency points. Abnormal waveforms are identified based on this peak distribution by calculating its variance. If an abnormal waveform exceeds a preset threshold, an adaptive filtering algorithm (LSMA) is applied to it, resulting in a filtered waveform sequence. The LMA uses the abnormal waveform as input and outputs a signal that minimizes the error. Amplitude and phase information are extracted from the filtered waveform sequence to determine the filtered frequency feature vector, obtained by separating the real and imaginary components. Noise residue is checked against the filtered frequency feature vector by calculating the energy proportion of non-dominant frequencies. If the noise residue exceeds a preset threshold, the frequency feature vector is optimized by iteratively adjusting the filtering parameters based on gradient descent to obtain the final filtered frequency feature vector.

[0030] Specifically, in one implementation, the mixed audio signal stream is first preprocessed to extract its frequency waveform features.

[0031] Specifically, the input mixed audio signal stream is converted into a time-frequency domain representation, for example, by decomposing the signal into multiple frequency components using a short-time Fourier transform, thereby obtaining the amplitude and phase information of the waveform. This extraction process helps to capture key patterns in the audio, such as the fundamental frequency in a speech signal or peaks in noise interference. Based on the extracted frequency waveform features, anomalous waveforms are further identified. Anomalous waveforms typically refer to those that deviate from normal audio patterns, such as bursts of noise or distorted signals.

[0032] In one possible implementation, a threshold comparison method is used to compare the amplitude value of a feature with a preset threshold. If the amplitude exceeds the threshold, it is marked as an anomaly. This identification relies on statistical analysis, such as calculating the mean and variance of the feature, to determine the boundaries of the anomaly, thereby ensuring the accuracy of the identification. After identifying the anomalous waveform, it is adaptively filtered. Adaptive filtering eliminates the anomaly by dynamically adjusting the filter parameters, for example, using the least mean square algorithm. The filter updates its weights in real time based on the error feedback of the input signal. In audio processing scenarios, this filtering can be applied to speech enhancement, such as removing background noise in telephone communication. The filtering process includes initializing the filter coefficients, calculating the error signal, and iteratively updating it to make the output signal closer to the desired clean waveform. The key to this step is the adaptive mechanism, which allows the filter to automatically optimize according to different audio environments, such as increasing the filter strength in high-noise environments.

[0033] For example, in a speech recognition system, the filtered frequency feature vector is determined by vector quantization of the filtered output. Specifically, the filtered waveform is converted into a fixed-dimensional vector, for example, using Mel-frequency cepstral coefficients, to facilitate subsequent processing such as pattern matching.

[0034] It should be noted that this method can be extended to various audio scenarios, such as noise removal of music signals.

[0035] In one embodiment, for real-time audio streams, feature extraction is combined with window function processing to reduce edge artifacts and ensure feature continuity. Furthermore, the adaptive filtering implementation can introduce a reference signal, such as noise captured by an auxiliary microphone, to enhance the filtering effect. This approach is particularly effective in conference recording scenarios, significantly improving audio clarity.

[0036] Preferably, when determining the filtered frequency feature vector, a normalization step can be added to map the vector values ​​to a uniform range to improve compatibility. For example, in audio monitoring applications, such vectors are used for anomaly detection. In another implementation, anomaly recognition is combined with machine learning models, such as training a classifier to learn the boundaries between normal and abnormal based on historical data, thereby improving the robustness of recognition.

[0037] Step S103: Determine the waveform fluctuation amplitude through the filtered frequency feature vector. If the waveform fluctuation amplitude exceeds the preset fluctuation threshold, it is identified as a jitter-related frequency band and enhanced to obtain an enhanced waveform stability representation.

[0038] The waveform fluctuation amplitude is obtained by calculating the square root of the sum of squared differences between vector elements after filtering. If the waveform fluctuation amplitude exceeds a preset fluctuation threshold, a tremor-related frequency band is identified. Fourier transform is used to calculate the spectrum of the tremor-related frequency band, and specific frequency components are amplified for enhancement, resulting in enhanced frequency band data. The enhanced frequency band data is then fused with the speech signal obtained from the filtered vector and smoothed to determine the enhanced stable waveform representation.

[0039] Specifically, in one implementation, the waveform fluctuation amplitude is determined by the filtered frequency feature vector. First, it's necessary to understand the concept of a frequency feature vector. This vector is typically a frequency domain representation extracted from the original waveform signal; after filtering to remove noise, the key frequency components are retained.

[0040] Specifically, fluctuation amplitude can be viewed as the degree of amplitude change at each frequency point in a vector, quantified, for example, by calculating the standard deviation or peak value difference of the vector elements. This method helps capture the dynamic characteristics of the waveform. This determination process provides a reliable basis for subsequent judgments, ensuring the accuracy of waveform analysis. Furthermore, if the waveform fluctuation amplitude exceeds a preset fluctuation threshold, it is identified as a jitter-related frequency band. Here, the preset fluctuation threshold is a value set based on experience or statistical data to distinguish between normal fluctuations and abnormal jitter. The identification process involves comparing the calculated amplitude with the threshold; if it exceeds, the corresponding frequency band is marked as jitter-related. This identification helps isolate interfering frequency bands, enabling targeted processing.

[0041] Preferably, enhancement is performed to obtain an enhanced waveform stability representation. A specific method of enhancement could be to amplify the amplitude of the identified frequency band, for example, by multiplying it by a coefficient greater than 1 to increase its weight in the overall vector, thereby making the jitter feature more prominent.

[0042] In one possible implementation, this enhancement employs filter adjustment, such as applying a high-pass filter to enhance the high-frequency jitter component while suppressing low-frequency noise, ultimately resulting in a more stable waveform representation for subsequent analysis.

[0043] It should be noted that there are several variations in determining the fluctuation range.

[0044] In one embodiment, instead of a simple standard deviation, the amplitude is calculated using power spectral density, which involves averaging the sum of squares of a vector to quantify the energy distribution. Specifically, the identification of the jitter-related frequency bands is not fixed but can be adjusted based on the application.

[0045] Understandably, the effect of the enhancement process is to improve the robustness of the signal.

[0046] For example, enhanced waveform representation can still highlight jitter characteristics in noisy environments, thereby improving detection accuracy. This objective description highlights the practical value of this application without introducing subjective evaluation.

[0047] In one embodiment, machine learning is used to assist in threshold setting, for example, by training a model using historical data to dynamically adjust the threshold, and then the magnitude is determined based on the filter vector and enhancement is identified, thereby obtaining an enhanced representation.

[0048] Step S104: Identify syllable interval duration by analyzing the enhanced waveform stability representation of the target time period, analyze the degree of interval shortening by analyzing the syllable interval duration, identify rapid speech rate characteristics, determine the time domain stretching compensation value by analyzing the rapid speech rate characteristics, and adjust the syllable interval duration according to the time domain stretching compensation value to obtain the time domain adjusted syllable sequence.

[0049] The syllable interval duration is calculated using enhanced waveform stabilization representation of the target time period. Interval shortening analysis is used to determine the degree of interval shortening, yielding a rapid speech rate characteristic. A temporal stretching compensation value is calculated for this rapid speech rate characteristic, and the syllable interval duration is adjusted based on this value to determine the temporally adjusted syllable sequence. Speech rate equalization data is obtained from the temporally adjusted syllable sequence. If the speech rate equalization data exceeds a preset threshold, the target time period is fused and processed to generate a balanced syllable sequence.

[0050] Specifically, in one implementation, the speech signal for the target time period is first preprocessed to obtain an enhanced waveform stability representation.

[0051] Specifically, the enhanced waveform stabilization representation is a stable signal form obtained by applying filtering and normalization processing to the original speech waveform.

[0052] For example, a low-pass filter is first used to remove high-frequency noise, and then energy normalization is used to ensure the stability of the waveform amplitude. This representation helps to highlight syllable boundaries and avoid waveform fluctuations caused by environmental interference. By analyzing this enhanced representation, the duration of syllable intervals can be identified.

[0053] For example, the positions of peaks and troughs in a waveform are detected, and the time difference between adjacent peaks is calculated as the interval duration. This process is particularly useful in speech recognition systems, enabling accurate capture of the speaker's rhythmic changes in real-time call scenarios. Furthermore, based on the identified syllable interval duration, the degree of interval shortening is analyzed to identify rapid speech rate characteristics. The degree of interval shortening is quantified by comparing the current interval duration with the average interval value at a standard speech rate.

[0054] For example, if the average duration of the intervals between multiple consecutive syllables is shortened by more than 20%, it is considered a rapid speech rate characteristic. This analysis process involves statistical methods, such as calculating the standard deviation and mean deviation of the interval durations, to identify patterns of accelerated speech rate.

[0055] In one possible implementation, for speech processing of conference recordings, this feature recognition can help distinguish between normal conversation and rapid debate segments, ensuring the accuracy of subsequent processing.

[0056] It should be noted that the identification of rapid speech rate features is not limited to a single threshold, but can also be combined with context, such as the speaker's emotional state, to further refine the judgment.

[0057] For example, the temporal stretching compensation value is determined by analyzing the characteristics of rapid speech rate. The core of this step is to calculate the compensation factor based on the degree of shortening.

[0058] For example, if the interval is shortened by 30%, the compensation value can be set to the corresponding stretching ratio to restore a natural speaking speed. The principle for determining the compensation value is to use a linear mapping relationship to map the feature intensity to the stretching coefficient.

[0059] For example, the higher the feature strength, the larger the stretch compensation value, thus achieving a balance in temporal adjustment. In speech synthesis applications, this compensation value can be applied to broadcasting systems to ensure that rapid input is adjusted to clear output. The key to this process is dynamic compensation, which can extend the interval without changing the pitch, resulting in a more natural listening experience.

[0060] Preferably, the syllable interval duration is adjusted according to the temporal stretching compensation value to obtain a temporally adjusted syllable sequence. Specifically, the adjustment process is achieved by inserting virtual frames or extending existing frames. For example, the shortened interval is temporally stretched by applying the compensation value to make it closer to the standard duration, thereby forming the adjusted sequence.

[0061] In one embodiment, for audio courseware on an online education platform, this adjustment can improve learners' comprehension and avoid information loss due to the lecturer's rapid speaking speed. Furthermore, this method is also applicable in different scenarios such as customer service recording analysis, ensuring sequence coherence through multiple iterative adjustments. In another embodiment, the entire process can be integrated into a mobile device's voice assistant. First, the user-inputted target time period is acquired, and enhanced waveform stabilization representation analysis is performed. Then, interval duration and abrupt features are identified step-by-step, and after determining the compensation value, the adjusted sequence is output. This integration demonstrates the versatility of this application.

[0062] For example, in navigation voice guidance, adjusting urgent commands to a standard speaking speed improves the user experience.

[0063] Specifically, the effect of the above adjustment process is to improve the robustness of speech processing.

[0064] For example, in noisy environments, compensating for abrupt features can reduce the recognition error rate. This objective effect has been experimentally verified, achieving improved accuracy in interval recovery across various speech samples. Further extending this approach to broadcast audio editing scenarios, the recorded target segment is first waveform-enhanced, then the degree of interval shortening is analyzed, features are identified, compensation values ​​are calculated, and the sequence is adjusted to obtain a smooth output. This scenario emphasizes the flexibility of the technology, enabling it to handle speech from different languages.

[0065] In one embodiment, considering multi-speaker environments, the syllable intervals of each speaker are analyzed separately to determine personalized compensation values. The adjusted sequences are then merged to ensure overall consistency. For example, in a teleconferencing system, this method can be applied in real time, analyzing the enhanced waveform to identify abrupt segments and adjusting the output to produce a clear sequence through compensation.

[0066] Step S105: Analyze the surrounding environmental audio data, identify the spectral characteristics of the environmental background audio, combine the time-domain adjusted syllable sequence, eliminate environmental interference, process the remaining noise residuals, and determine the pure candidate wake words.

[0067] The system acquires ambient audio data, identifies the spectral characteristics of the background audio, and combines this with a time-domain adjusted syllable sequence to eliminate environmental interference, resulting in a preliminary purified sequence. For this preliminary purified sequence, residual noise is processed by converting the time-domain signal to a frequency-domain representation using Fourier transform, calculating the frequency-domain residual value, determining the residual suppression threshold, and adjusting the sequence to obtain an intermediate purified sequence. Candidate wake-up word features are extracted from the intermediate purified sequence. If the features match a preset threshold, multi-channel audio data is fused to generate an extended sequence, resulting in an enhanced candidate sequence. The purity of the enhanced candidate sequence is determined by using a Gaussian mixture model with sequence features as input to evaluate word boundaries and output boundary positions, thus identifying pure candidate wake-up words.

[0068] Specifically, in one implementation, the process of analyzing ambient audio data begins with acquiring real-time audio signals via a microphone array. This acquisition is targeted at smart devices such as smart speakers or voice assistants, ensuring that a mixture of user voice and ambient noise is captured.

[0069] Specifically, the audio data is digitized at a sampling rate to form a raw waveform sequence. This step aims to provide foundational data for subsequent feature extraction, enabling effective capture of sound variations in a home or office environment. In this way, the system can initially distinguish between foreground speech and background noise, achieving a comprehensive recording of environmental audio. Further, identifying the spectral characteristics of the environmental background audio involves applying a Fast Fourier Transform to the acquired audio data, converting the time-domain signal into a frequency-domain representation.

[0070] In one possible implementation, the audio frames are first weighted using a window function, such as a Hamming window, to reduce spectral leakage. Then, the power spectral density of each frame is calculated to extract the spectral peaks and distribution characteristics of the background noise. These spectral characteristics can be represented as a vector containing energy values ​​across multiple frequency bands. This identification process helps quantify the type of environmental interference, such as wind noise or traffic noise, and when applied to smart home devices, it can improve the accuracy of background modeling.

[0071] It should be noted that combining the time-domain adjusted syllable sequence is one of the core steps. Here, the syllable sequence refers to the sequence of speech units segmented from the original audio, which is aligned with uneven pronunciation speeds through time-domain adjustments such as dynamic time warping.

[0072] Specifically, the Mel-frequency cepstral coefficients are first extracted as syllable features. Then, a temporal adjustment algorithm is applied to fuse these sequences with previously identified spectral features. For example, in a voice assistant wake-up scenario, if a user speaks in a noisy kitchen environment, the system adjusts the duration of the syllable sequence to align it with the background spectrum, thereby generating a comprehensive acoustic representation vector. This fusion process, achieved through vector concatenation or weighted summation, effectively captures rhythmic variations in speech and the spectral impact of noise, ensuring more robust input data for subsequent processing. When processing variable-speed speech, temporal adjustment compensates for timing deviations caused by environmental factors, thus providing a more stable feature base.

[0073] Preferably, environmental interference is eliminated through an adaptive filter.

[0074] In one embodiment, a noise model is constructed based on the identified spectral features, and then the estimated interference components are subtracted from the mixed audio.

[0075] For example, in smart speaker applications, if there is background air conditioning noise, the system calculates the noise's spectral template and eliminates the corresponding portion in real time, thus preserving the main speech signal. This elimination process can use the minimum mean square error criterion to optimize the filtering parameters, ensuring the accuracy of interference removal.

[0076] In one implementation, processing residual noise involves further analysis of the canceled signal. Specifically, residual noise refers to subtle noise components that remain after interference removal, and is processed through residual modeling. First, the residual signal, i.e., the difference between the processed audio and the ideal clean signal, is calculated, and then methods such as Wiener filtering are applied to suppress these residuals.

[0077] For example, in in-vehicle voice systems, if road noise remains, the system estimates the statistical characteristics of the residuals, such as the mean and variance, and minimizes the residual energy by iteratively updating the filter parameters. The principle behind this process is that residuals often exhibit a random distribution; by modeling their probability density function, targeted suppression strategies can be designed to obtain clearer speech signals. In practical applications, this processing can significantly improve the accuracy of wake-word detection in high-noise environments because it specifically addresses the fine-grained noise issues missed by traditional filtering. This technical effect is reflected in the system's adaptability to complex environments, making wake-up responses more reliable.

[0078] For example, in an office environment, the system first acquires audio data including conversations and keyboard tapping, then identifies background spectral features such as low-frequency tapping peaks. After combining this with a time-domain adjusted syllable sequence, interference is eliminated, residuals are processed, and a clean wake-up word is ultimately determined. Further, clean candidate wake-up words are matched based on the processed signal.

[0079] In one possible implementation, a Hidden Markov Model is used to model the clean signal and calculate a similarity score with a predefined wake word template.

[0080] Specifically, if the score exceeds a threshold, it is selected as a candidate word. For example, in smart home devices, this selection process can handle multiple candidates, such as "Hey Assistant," ensuring accurate recognition even in noisy environments.

[0081] Understandably, in another embodiment, the entire process is applied to a mobile phone voice assistant. First, ambient audio, such as street noise, is analyzed to identify spectral characteristics. Then, the syllable sequence is adjusted, interference is eliminated, and residuals are processed, ultimately outputting a clean wake word. This implementation demonstrates the versatility of this application on mobile devices.

[0082] Preferably, the above method can be implemented using an embedded processor, adjusting parameters in different smart devices to adapt to specific scenarios, thereby covering various voice interaction environments such as home, office, and in-vehicle.

[0083] Step S106: Perform similarity matching between the clean candidate wake words and the preset emergency wake word template to determine the intent recognition result, specifically: Obtain pure candidate wake words and extract Mel frequency cepstral coefficients from them as candidate vectors; Randomly select an emergency wake-up word template from the emergency wake-up word template library and extract the emergency word vector; Calculate the cosine similarity between the candidate vector and the emergency word vector; The cosine similarity is used for matching. If the similarity exceeds a set threshold, the match is successful and a preliminary intent category is obtained. Otherwise, another emergency wake-up word template is obtained, the cosine similarity is calculated, and a matching judgment is made until the obtained emergency wake-up word templates are repeated, thus obtaining a preliminary set of intent categories; The preliminary intent category set is fused and analyzed to determine the final intent recognition result.

[0084] Specifically, in one implementation, the system first preprocesses the input speech signal to extract clean candidate wake words.

[0085] Specifically, active speech segments are identified using a speech endpoint detection algorithm, and then noise suppression techniques are applied to remove background interference. For example, spectral subtraction is used to eliminate noise components from the spectrum, thereby obtaining clean speech segments. This purification process ensures that candidate wake words are not affected by environmental noise, improving the accuracy of subsequent matching.

[0086] For example, when a user issues a voice command in a noisy kitchen environment, this step effectively isolates clear word fragments, providing a reliable foundation for emergency intent recognition. In this way, the system optimizes the original speech, supporting applications with various home devices such as smart speakers. Further, the extracted clean candidate wake words are matched with preset emergency wake word templates based on similarity. This matching process is the core of this application, involving the calculation and comparison of feature vectors. First, Mel-frequency cepstral coefficients are extracted from the clean candidate wake words as candidate vectors, and then cosine similarity is calculated with emergency word vectors (such as "help" or "emergency") in the template library. The specific calculation process includes normalizing the two vectors and then taking their inner product. For example, if the candidate vector is A and the template vector is B, the similarity is A·B / (|A||B|).

[0087] It should be noted that this similarity matching quantifies the semantic proximity of speech, and in business applications, it is used in emergency response systems to ensure that only highly similar words trigger intent recognition. A key feature of this step is the integration of dynamic time warping technology to handle variations in pronunciation duration, thereby adapting to different users' speaking habits.

[0088] In one possible implementation, a similarity threshold of 0.8 or higher is considered a successful match. This helps voice assistant devices quickly distinguish between ordinary commands and emergency calls, avoiding false triggers. Through a detailed matching mechanism, the system improves the robustness of intent recognition and can effectively handle voice requests for help in real-world home environments, such as in cases of sudden emergencies involving the elderly, resulting in higher response efficiency.

[0089] For example, after the matching is completed, the system determines the intent recognition result based on the similarity result.

[0090] Specifically, if the similarity exceeds a preset threshold, it is determined to be an urgent intent, and the corresponding response module is activated; otherwise, it is considered non-urgent and ignored. This logical judgment ensures the accuracy of the system.

[0091] Preferably, in another implementation, similarity matching can be extended to multimodal fusion, such as combining speech rate and volume features to further verify intent. For example, when a candidate wake word has moderate similarity but a faster speech rate, the system can increase its urgency probability. This extension enhances the flexibility of this application and makes it suitable for diverse smart home scenarios.

[0092] Understandably, the matching process described above is based on statistical pattern recognition, ensuring high accuracy even in noisy environments. In business processes, this means the entire chain from voice input to intended output is optimized.

[0093] In one embodiment, the extraction of pure candidate wake words can be further refined using an adaptive filter.

[0094] Specifically, a cleaner word is obtained by estimating the noise power spectrum through Wiener filtering and subtracting it from the signal. This method can handle interference such as background noise from a television in home voice interaction, supporting the aforementioned purification features. Furthermore, the intent recognition result can be used to trigger an alarm system, such as automatically dialing an emergency number after a successful match. This application demonstrates the versatility of this application within the same field.

[0095] It should be noted that the construction of preset emergency wake-up word templates involves the collection of training datasets, such as extracting template vectors from historical emergency voice samples, to ensure the reliability of the matching. For example, in a smart home bedroom scenario, when a user utters a vague "help" word at night, the system quickly identifies the intent through similarity matching, activates the lights, and notifies family members.

[0096] Step S107: Analyze the confidence score of the intent recognition result. If the confidence score exceeds the response trigger condition, the emergency response assistance process is triggered, and an emergency response command execution signal is obtained, specifically: The confidence score is obtained based on the intent recognition result; The confidence score and the response triggering condition are calculated using a cosine similarity algorithm to obtain a score matching score; wherein, the response triggering condition is represented as a preset threshold vector; The score is matched with a score, and if the score exceeds the response triggering condition, the emergency response process is activated and the activation status of the process is determined. Based on the activation status of the process, a help request execution signal is obtained and integrated with the signal generation interface to obtain an emergency response instruction; After the emergency response command is generated, the emergency response command execution signal is obtained by adjusting the intent recognition result.

[0097] For example, the confidence score is calculated using a probability distribution model, which first extracts input features, such as keywords and context vectors, and then uses a softmax function to output the probability value of each intent category.

[0098] Understandably, the principle behind this analysis lies in the multi-classifier structure of machine learning, where the training data includes historical query samples, and the model optimizes parameters through gradient descent to improve accuracy.

[0099] In one embodiment, the system triggers an emergency response assistance process when the confidence score exceeds a preset response trigger condition. For example, once the confidence score exceeds the threshold, the system automatically generates an assistance signal and sends an alert to relevant personnel or external services. This mechanism enhances the system's robustness and enables timely responses to potential risks.

[0100] In another implementation, the confidence analysis for intent confirmation employs an ensemble learning framework, in which multiple sub-models jointly vote to calculate a score.

[0101] Specifically, the framework includes decision tree and neural network sub-modules. First, features, such as word embedding vectors, are extracted from the input. Then, each sub-module outputs preliminary probabilities, and the framework obtains the final confidence score through a weighted average. The principle behind this framework is to reduce the bias of a single model, effectively identify complex intentions, and thus improve the system's intelligence level.

[0102] It should be noted that the execution signal for the emergency response request process is obtained through a feedback loop mechanism. For example, after the system sends a request for assistance, it waits for a confirmation signal for the response instruction. This signal can be a digital code: 0 indicates execution, and 1 indicates delay. This process ensures the continuity of the response chain and avoids delays.

[0103] In one possible implementation, the calculation of the confidence score involves adjusting the posterior probability and updating the initial probability based on Bayesian principles.

[0104] Specifically, a prior distribution is first established, and then the posterior distribution is updated based on the results of new data calculations. The resulting score is used to confirm intent. This process enables more accurate risk prediction.

[0105] For example, the acquisition of an emergency response command execution signal may include multi-level verification. For instance, the system first internally verifies the score, then seeks external assistance, and finally executes an automated script upon receiving the signal. This design delivers a rapid response capability in business operations.

[0106] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for activating a smart microphone based on speech recognition, characterized in that: Includes the following steps: The system collects user voice signals and ambient audio data, captures multi-channel input signals using a microphone array, and obtains a mixed audio signal stream. It extracts frequency waveform features from the mixed audio signal stream and identifies abnormal waveforms, performing adaptive filtering on these abnormal waveforms to determine the filtered frequency feature vector. Based on the filtered frequency feature vector, it determines the waveform fluctuation amplitude. If the waveform fluctuation amplitude exceeds a preset fluctuation threshold, it is identified as a tremor-related frequency band and amplified to obtain an enhanced waveform stability representation. It analyzes the enhanced waveform stability representation of the target time period to identify syllable interval duration and analyzes the degree of interval shortening based on the syllable interval duration. It identifies and analyzes rapid speech rate features to determine a temporal stretching compensation value, adjusts the syllable interval duration based on the temporal stretching compensation value, and obtains a temporally adjusted syllable sequence. It analyzes ambient audio data, identifies the spectral features of the background audio, combines this with the temporally adjusted syllable sequence, eliminates environmental interference, processes residual noise, and determines clean candidate wake words. Finally, it performs similarity matching between the clean candidate wake words and a preset emergency wake word template to determine the intent recognition result. Analyze the confidence score of the intent recognition result. If the confidence score exceeds the response trigger condition, the emergency response assistance process is triggered, and an emergency response instruction execution signal is obtained.

2. The method for activating a smart microphone based on speech recognition according to claim 1, characterized in that: The process of collecting user voice signals and ambient audio data involves using a microphone array to capture multi-channel input signals, resulting in a mixed audio signal stream, including: Based on the ambient audio data, a beamforming algorithm is used to process the multi-channel input signals of the microphone array to obtain a directionally enhanced speech signal; Real-time timestamps are obtained for the targeted enhanced speech signal to obtain an annotated mixed audio signal stream; The ambient noise level is determined based on the labeled mixed audio signal stream. If it exceeds a preset threshold, noise suppression processing is applied to obtain a purified audio stream. The user's voice features are extracted from the purified audio stream to determine the final signal recognition result.

3. The method for activating a smart microphone based on speech recognition according to claim 1, characterized in that: The step of extracting the frequency waveform features of the mixed audio signal stream and identifying abnormal waveforms, performing adaptive filtering on the abnormal waveforms, and determining the filtered frequency feature vector includes: Frequency waveform features are obtained from the mixed audio signal stream based on Fourier transform, and the peak distribution in the frequency waveform features is determined. Based on the peak distribution, abnormal waveforms are identified. If the abnormal waveform exceeds a preset threshold, the least mean square algorithm is used to adaptively filter the abnormal waveform to obtain a filtered waveform sequence. Amplitude and phase information are extracted from the filtered waveform sequence to determine the filtered frequency feature vector; Noise residue is checked on the filtered frequency feature vector. If the noise residue is higher than the preset threshold, the filtered frequency feature vector is optimized by iteratively adjusting the filtering parameters to obtain the final optimized filtered frequency feature vector.

4. The method for activating a smart microphone based on speech recognition according to claim 1, characterized in that: The process of determining waveform fluctuation amplitude based on the filtered frequency feature vector, and identifying it as a jitter-related frequency band if the waveform fluctuation amplitude exceeds a preset fluctuation threshold, and then amplifying it to obtain an enhanced waveform stability representation, includes: The square root of the sum of squares of the differences between vector elements is calculated using the filtered frequency feature vector to obtain the waveform fluctuation amplitude and identify the jitter-related frequency bands. Fourier transform is used to calculate the spectrum of the aforementioned vibration-related frequency band, and specific frequency components are amplified to enhance the signal, resulting in enhanced frequency band data. The speech signal is obtained from the filtered frequency feature vector, fused with the enhanced frequency band data, and smoothed to determine the enhanced waveform stable representation.

5. The method for activating a smart microphone based on speech recognition according to claim 1, characterized in that: The enhanced waveform stabilization representation of the target time period is analyzed to identify syllable interval duration, and the degree of interval shortening is analyzed based on the syllable interval duration. Identify and analyze rapid speech rate characteristics to determine temporal stretching compensation values, and adjust the syllable interval duration according to the temporal stretching compensation values ​​to obtain a temporally adjusted syllable sequence, including: The syllable interval duration is calculated based on the enhanced waveform stability representation of the target time period, the degree of interval shortening is determined, and the rapid speech rate characteristics are obtained. Calculate a temporal stretching compensation value for the rapid speech rate feature, adjust the syllable interval duration based on the temporal stretching compensation value, and determine the temporally adjusted syllable sequence.

6. The method for activating a smart microphone based on speech recognition according to claim 1, characterized in that: The analysis of surrounding environmental audio data identifies the spectral characteristics of background audio, combines this with a time-domain adjusted syllable sequence, eliminates environmental interference and processes residual noise, and determines clean candidate wake words, including: The surrounding environmental audio data is acquired, the spectral characteristics of the environmental background audio are identified, and combined with the time-domain adjusted syllable sequence, environmental interference is eliminated and a preliminary purification sequence is obtained. For the remaining noise residuals in the preliminary purification sequence, the time-domain signal is converted into a frequency-domain representation according to the Fourier transform and the frequency-domain residual value is calculated. The residual suppression threshold is determined, and the syllable sequence is adjusted to obtain the intermediate purification sequence. Candidate wake word features are extracted from the intermediate cleaned sequence. If the candidate wake word features match a preset threshold, the multi-channel audio data are fused to generate an extended sequence, thus obtaining an enhanced candidate sequence. The purity of the enhanced candidate sequence is determined by using a Gaussian mixture model with the enhanced candidate sequence features as input to evaluate the boundary of the word and output the boundary position to determine the pure candidate wake word.

7. The method for activating a smart microphone based on speech recognition according to claim 1, characterized in that: The step of performing similarity matching between the purified candidate wake words and the preset emergency wake word template to determine the intent recognition result includes: Obtain pure candidate wake words and extract Mel frequency cepstral coefficients from them as candidate vectors; Randomly select an emergency wake-up word template from the emergency wake-up word template library and extract the emergency word vector; Calculate the cosine similarity between the candidate vector and the emergency word vector; The cosine similarity is used for matching. If the similarity exceeds a set threshold, the match is successful and a preliminary intent category is obtained. Otherwise, another emergency wake-up word template is obtained, the cosine similarity is calculated, and a matching judgment is made until the obtained emergency wake-up word templates are repeated, thus obtaining a preliminary set of intent categories; The preliminary intent category set is fused and analyzed to determine the final intent recognition result.

8. The method for activating a smart microphone based on speech recognition according to claim 1, characterized in that: The analysis of the confidence score of the intent recognition result, if the confidence score exceeds the response trigger condition, triggers the emergency response assistance process and obtains an emergency response command execution signal, including: The confidence score is obtained based on the intent recognition result; The confidence score and the response triggering condition are calculated using a cosine similarity algorithm to obtain a score matching score; wherein, the response triggering condition is represented as a preset threshold vector; The score is matched with the score. If the score exceeds the response triggering condition, the emergency response help process is activated and the activation status of the process is determined. Based on the activation status of the process, a help request execution signal is obtained and integrated with the signal generation interface to obtain an emergency response instruction; After the emergency response command is generated, the emergency response command execution signal is obtained by adjusting the intent recognition result.

Images