An audio stream control method based on voice interaction

Abstract

This invention relates to the field of voice interaction and audio stream processing technology, specifically to an audio stream control method based on voice interaction. It employs a microphone array composed of 3-6 piezoelectric pickup units to synchronously acquire user voice signals and environmental noise signals in real time. The signal amplitude is normalized to 0.1V-0.5V, and then preliminary noise reduction is performed using an adaptive interference cancellation algorithm to remove some fixed-frequency noise. This invention combines dynamic threshold configuration with adaptive noise reduction to transform environmental noise from an interference source into a signal source that cancels out interference, while simultaneously recovering noise vibration energy. This results in higher recognition accuracy and a lower false trigger rate, effectively balancing the high sensitivity and anti-interference capabilities of voice interaction, and solving the problem of a sharp drop in recognition accuracy in noisy scenes caused by fixed thresholds in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of voice interaction and audio stream processing technology, specifically to an audio stream control method that achieves adaptive noise reduction, accurate intent recognition, and low false triggering through voice commands, applicable to scenarios such as smart homes, in-vehicle audio-visual systems, and wearable audio devices. Background Technology

[0002] Voice interaction technology is now widely used in the field of audio stream control, becoming the core interaction method for devices such as smart speakers, in-vehicle multimedia, and TWS earphones. Through voice commands, audio stream operations such as play, pause, skip tracks, and volume adjustment can be achieved, improving the convenience of user interaction.

[0003] Existing audio stream voice control solutions mostly adopt a single microphone pickup + fixed threshold recognition architecture. Some solutions introduce basic noise reduction algorithms and use cloud or local models to complete speech-to-text conversion and command parsing, and then send control commands to the audio playback unit. High-end solutions are supplemented with voiceprint recognition for user verification, but the overall system still relies on a single voice signal as the core control basis.

[0004] The existing technology has the following drawbacks: 1. Environmental noise can easily lead to false triggering. Fixed recognition thresholds are used in quiet and high-noise scenes. Under strong noise, the recognition accuracy drops sharply. 2. It cannot accurately distinguish the intent of commands, which can easily lead to confusion in the control of multiple audio streams such as music / video; 3. The contradiction between response speed and system complexity is prominent, it is difficult to balance high sensitivity and anti-interference, and there is a lack of multi-modal verification mechanism, resulting in insufficient security and control accuracy. 4. Single auditory feedback makes operation confirmation difficult in screenless scenarios, resulting in a poor user experience. Summary of the Invention

[0005] In order to overcome the above-mentioned technical problems, the purpose of this invention is to provide an audio flow control method based on voice interaction: to solve the problems mentioned in the background art.

[0006] The objective of this invention can be achieved through the following technical solutions: An audio stream control method based on voice interaction includes the following steps: S1. A microphone array consisting of 3-6 piezoelectric pickup units is used to synchronously acquire user voice signals and environmental noise signals in real time. The acquisition frequency is set to 16kHz-48kHz, and the sampling accuracy is 16bit-24bit. Automatic gain control (AGC) is performed on the acquired mixed signal to normalize the signal amplitude to 0.1V-0.5V. Then, an adaptive interference cancellation algorithm is used for preliminary noise reduction to remove some fixed frequency noise. S2. The ambient noise decibel value is detected in real time by a noise sensor at a frequency of 1 time / 100ms. A corresponding mapping relationship between noise decibel and speech recognition trigger threshold is established. The trigger threshold is adaptively adjusted according to the real-time detected noise decibel. At the same time, the RNNoise deep learning noise reduction model is called to perform time-domain-frequency domain joint filtering on the preprocessed signal to remove non-human low-frequency noise (50Hz-200Hz) and retain effective speech signal (300Hz-3400Hz). S3. Input the noise-reduced speech signal into the local lightweight speech recognition model (model parameters not exceeding 500M, inference latency ≤50ms) to complete the speech-to-text operation. Extract keywords and perform semantic analysis on the recognized text, match it with the preset standard command library (including commands such as play, pause, skip songs, volume adjustment, fast forward, rewind, and switch audio sources), and read the scene tags of the audio stream currently running on the device (music, video, podcast, audiobook, etc.). Combine the scene tags and semantic analysis results to accurately distinguish the command control object and generate standardized control commands. S4. Extract voiceprint features from the identified speech signal. The extracted voiceprint features include 12-16 dimensional feature parameters such as fundamental frequency, spectral envelope, and Mel-frequency cepstral coefficients (MFCC). Compare these features with the user's voiceprint key pre-stored on the device. If the similarity is ≥90%, the user is deemed legitimate. Based on this, a secondary verification is performed using the user's biometrics. The secondary verification uses either laryngeal electromyography (EMG) signal verification or lip movement feature verification. After the verification is successful, standardized control commands are sent to the audio playback unit to execute the corresponding audio flow control operation. S5. After the command is executed, auditory sound feedback and vibration tactile feedback are output simultaneously. The auditory sound feedback uses emotional sound effects, and different control commands correspond to different sound effect parameters. The vibration tactile feedback is achieved through the device's built-in vibration unit, and different control commands correspond to different vibration modes, ensuring that users can quickly confirm the operation results in screenless scenarios.

[0007] As a further aspect of the present invention, step S1 specifically includes the following sub-steps: S11. The piezoelectric pickup unit of the microphone array is made of lead zirconate titanate (PZT) piezoelectric material, with a pickup sensitivity of -38dB±2dB and a frequency response range of 50Hz-20kHz. While acquiring the user's voice signal, it simultaneously acquires the mechanical vibration signal generated by the ambient noise. S12. The collected noise vibration signal is converted into DC power through the piezoelectric effect with a conversion efficiency of 15%-25%. The converted power is stored in a micro energy storage unit (capacity of 100mAh-500mAh) for standby power supply of the microphone array. S13. The converted noise vibration electrical signal is filtered and amplified, with the amplification factor set to 20dB-40dB. The processed signal is analyzed by the adaptive least mean square (LMS) algorithm to generate an inverse filter signal with the same amplitude and opposite phase as the noise signal. S14. The reverse filtered signal is superimposed onto the original acquired mixed speech and noise signal to achieve active noise reduction, with a noise reduction amount of 10dB-18dB. At the same time, the superimposed signal is de-aliased to avoid signal distortion.

[0008] As a further aspect of the present invention, step S2 specifically includes the following sub-steps: S21. The noise sensor uses an electret microphone sensor with a measurement range of 30dB-120dB and a measurement accuracy of ±1dB. It detects the ambient noise decibel value in real time and updates the detection data every 100ms. S22. Establish a decibel-threshold mapping table, with the following specific mapping relationships: When the noise level is 0-40dB (quiet scene), the speech recognition trigger threshold is set to 0.3 (normalized amplitude); when the noise level is 40-60dB (everyday scene), the trigger threshold is set to 0.5 (normalized amplitude); when the noise level is 60-85dB (noisy scene), the trigger threshold is set to 0.7 (normalized amplitude); when the noise level is above 85dB (strong noise scene), the trigger threshold is set to 0.8 (normalized amplitude), and the ultrasonic active anti-interference mechanism is activated simultaneously. S23. Call the RNNoise deep learning model, which adopts a recurrent neural network (RNN) architecture and includes an input layer, hidden layers (3-5 layers), and an output layer. The input layer receives the preprocessed speech signal, the hidden layer extracts features through the activation function (ReLU), and the output layer outputs the denoised clean speech signal, with a filtering efficiency of ≥90% for non-human voice low-frequency noise in the 50Hz-200Hz range. S24. Perform endpoint detection on the denoised speech signal, and use a combination of short-time energy and short-time zero-crossing rate to accurately distinguish between speech segments and silence segments, remove silence segment signals, reduce invalid data processing, and improve system response speed.

[0009] As a further aspect of the present invention, step S3 specifically includes the following sub-steps: S31. The local lightweight speech recognition model adopts an architecture that combines deep neural networks (DNN) and hidden Markov models (HMM). The model training dataset contains more than 100,000 speech command data with different accents and different noise environments, with a recognition accuracy of ≥95% (quiet scene). S32. Extract keywords from the speech-to-text recognition results. Use the TF-IDF algorithm to extract keywords with an extraction threshold of 0.6. Match the extracted keywords with the standard instruction library. If the similarity is ≥85%, it is considered a valid instruction. S33. The device records the currently running audio stream information in real time and assigns a unique scene tag to each audio stream. Scene tags include music, video, podcast, and audiobook. The tag information is stored in the scene management module on the device and can be updated in real time. S34. Combining keyword matching results with scene tags, perform semantic intent parsing. When the keyword "pause" is identified and the scene tag is "music", generate a standardized control command to "pause music"; when the scene tag is "video", generate a standardized control command to "pause video" to avoid confusion in multi-audio stream control.

[0010] As a further aspect of the present invention, step S4 specifically includes the following sub-steps: S41. Voiceprint feature extraction adopts the Mel frequency cepstral coefficient (MFCC) method to extract 16-dimensional feature parameters. The sampling window length is set to 25ms and the frame shift is set to 10ms. The extracted feature parameters are normalized to remove the influence of individual differences. S42. Voiceprint comparison uses the Dynamic Time Warping (DTW) algorithm, with a comparison threshold of 0.1. When the similarity is ≥90%, the user is considered legitimate; if the similarity is <90%, the control command is refused and a prompt sound effect is output. S43. When the secondary verification uses laryngeal electromyography (EMG) signals, the device's built-in EMG sensor collects the micro-vibration signals of the laryngeal muscles when the user speaks. The collection frequency is 1kHz. The peak value and effective value features of the EMG signals are extracted and compared with the pre-stored EMG feature templates. If the similarity is ≥88%, the verification passes. S44. When the secondary verification uses lip movement feature verification, the device camera captures the lip movement image of the user speaking in real time at a frame rate of 15fps. The lip movement image is preprocessed (grayscale, noise reduction, normalization), and lip movement feature points (12-16) are extracted and matched with the pre-stored lip movement mouth shape library. If the match is successful, the verification is passed. S45. Only when both voiceprint verification and secondary verification pass will the standardized control command be sent to the audio playback unit to execute audio flow control operation. If the verification fails, the operation will be terminated to ensure control security.

[0011] As a further aspect of the present invention, step S5 specifically includes the following sub-steps: S51. The specific parameters for auditory sound effect feedback are as follows: the play command corresponds to an upward pitch, with the frequency rising from 440Hz to 880Hz and a duration of 300ms; the pause command corresponds to a fade-out sound effect, with the frequency falling from 880Hz to 440Hz and a duration of 300ms; the skip command corresponds to a short prompt tone, with a frequency of 660Hz and a duration of 150ms; the volume adjustment command corresponds to a continuous sound effect, with a frequency of 550Hz and a duration synchronized with the adjustment duration. S52. Vibration tactile feedback is achieved through an eccentric motor vibration unit, with a vibration frequency set to 100Hz-200Hz and a vibration amplitude set to 0.5g-1g. Specifically, the pause command corresponds to one short vibration with a duration of 100ms; the song skip command corresponds to two short vibrations, each with a duration of 100ms and an interval of 50ms; and the volume adjustment command corresponds to a long vibration with a duration consistent with the adjustment operation duration (not exceeding 2s). S53. Sound and vibration feedback are activated simultaneously with a startup delay of ≤50ms, ensuring that users can receive both auditory and tactile feedback at the same time, thus improving the accuracy of operation confirmation in screenless scenarios.

[0012] As a further aspect of the present invention: the micro energy storage unit uses a lithium polymer battery with a charge-discharge cycle count of ≥1000 times and an operating temperature range of -10℃ to 60℃. When the energy storage unit's charge is ≥80%, noise and vibration energy recovery is stopped; when the charge is ≤20%, the energy recovery priority enhancement mode is activated, and the energy recovery efficiency is increased to 30% to maintain the microphone array's standby power supply, which can reduce the device's standby power consumption by 40%-42%.

[0013] As a further aspect of this invention: In an 85dB high noise environment, after dynamic threshold adjustment and RNNoise model noise reduction, the speech recognition accuracy is ≥88%, and the false trigger rate is reduced by ≥30% compared to the traditional fixed threshold scheme; In a 100dB strong noise environment, the ultrasonic active anti-interference mechanism is activated to emit 25-40kHz ultrasonic waves to form an acoustic barrier, which can maintain the speech recognition accuracy at over 75%.

[0014] As a further aspect of the present invention: after dual modal verification of voiceprint features and biometric features, the command error rate is reduced by ≥90%, which can effectively prevent others from impersonating voice commands for control. It is suitable for high-security control scenarios such as vehicle payment and smart home access control, and meets financial-grade security requirements.

[0015] As a further aspect of the present invention: by adopting a combination of auditory sound effects and tactile vibration feedback, the accuracy of operation confirmation in screenless scenarios (driving, sports) is improved by ≥50%, and compared with single auditory feedback, user interaction satisfaction is improved by ≥35%. Psychological experiments have verified that the user operation error rate is reduced by more than 60%.

[0016] The beneficial effects of this invention are: 1. By combining dynamic threshold configuration with adaptive noise reduction, environmental noise is transformed from an interference source into a signal source that cancels out the interference. At the same time, the noise vibration energy is recovered, resulting in higher recognition accuracy and lower false trigger rate. It effectively balances the high sensitivity and anti-interference of voice interaction and solves the problem of sharp drop in recognition accuracy in noisy scenes caused by the fixed threshold of existing technologies. 2. By introducing a combination of audio stream scene tags and semantic analysis, different control objects such as music, video, and podcasts can be accurately distinguished, completely solving the problem of multi-audio stream command confusion. The accuracy of command parsing is greatly improved, avoiding the intention misjudgment caused by single speech recognition in existing technologies. 3. By adopting a local lightweight speech recognition model and combining it with a multimodal verification mechanism, the response speed is faster without significantly increasing the system complexity. At the same time, the dual verification of voiceprint and biometric features reduces the false execution rate, thus resolving the contradiction between the response speed, system complexity, anti-interference ability and security of existing technologies. 4. By combining auditory sound effects with vibration and tactile feedback, the feedback modes corresponding to different commands are clearly defined. Users can quickly confirm the operation results without looking at the screen. The accuracy of operation is improved in screenless scenarios such as driving and sports, and user satisfaction is improved. This improves the shortcomings of the single auditory feedback in the existing technology. 5. By converting noise vibration energy into electrical energy and recycling it through a piezoelectric pickup unit, the microphone array is powered in standby mode, reducing the device's standby power consumption and effectively extending the battery life of portable audio devices (such as TWS earphones and smart speakers), thus solving the problem of excessive power consumption in existing voice control devices. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of an audio stream control method based on voice interaction in this invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to specific embodiments. These embodiments are intended to fully disclose the technical solutions of the present invention, enabling those skilled in the art to implement the present invention, and are not intended to limit the scope of protection of the present invention.

[0020] Example 1: In-vehicle multimedia audio stream control 1. Hardware configuration: vehicle microphone array (composed of 4 lead zirconate titanate piezoelectric pickup units, pickup sensitivity -38dB, frequency response 50Hz-20kHz), electret noise sensor (measurement range 30dB-120dB, accuracy ±1dB), vehicle audio processor (1.2GHz main frequency, 1GB memory), vibration speaker (vibration frequency 150Hz, amplitude 0.8g), laryngeal electromyography sensor (sampling frequency 1kHz), miniature lithium polymer energy storage unit (300mAh, 1000 charge-discharge cycles); 2. Steps to be followed: S1. The microphone array simultaneously acquires the driver's voice signal and the engine and wind noise signals inside the vehicle at a sampling frequency of 32kHz and a sampling precision of 24bit. The signal amplitude is normalized to 0.3V through AGC, and the piezoelectric unit converts the noise vibration into electrical energy (conversion efficiency of 20%), which is stored in the energy storage unit. At the same time, the LMS algorithm generates an inverse filter signal, which is superimposed to achieve active noise reduction, with a noise reduction of up to 15dB. S2. The noise sensor detects in-vehicle noise every 100ms, with a noise range of 65-75dB. Based on the mapping table, the voice recognition trigger threshold is automatically set to 0.7. The RNNoise deep learning model is called to filter low-frequency engine noise from 50Hz to 200Hz with a filtering efficiency of 92%. Then, short-time energy and zero-crossing rate detection are used to remove silent signals. S3. Input the noise-reduced speech signal into the local lightweight speech recognition model (parameters 450M, inference latency 45ms) to complete speech-to-text conversion. When "pause music" is recognized, extract the keywords "pause" and "music" through the TF-IDF algorithm, match the standard instruction library, read the current scene label as "music", and generate the standardized control instruction "pause music". S4. Extract the 16-dimensional MFCC voiceprint features of the speech, compare them with the pre-stored driver voiceprint key using the DTW algorithm, and the similarity is 95%, thus determining that the user is legitimate; collect the driver's laryngeal electromyography signal using an electromyography sensor, extract the peak and effective value features, compare them with the pre-stored template, and the similarity is 90%, thus passing the second verification; send the control command to the vehicle audio playback unit to execute the pause operation. S5. Synchronously outputs a gradually decreasing sound effect (frequency drops from 880Hz to 440Hz, duration 300ms) and a short vibration (150Hz, 100ms), allowing the driver to confirm the operation is complete without looking at objects; 3. Performance data: 92% recognition accuracy, 5% false trigger rate, 98% operation accuracy, 42% reduction in standby power consumption, and 0.1s response time in an 85dB noise environment, meeting the usage requirements of in-vehicle scenarios.

[0021] Example 2: Smart Speaker Home Control 1. Hardware configuration: Dual microphone array (piezoelectric pickup unit, sensitivity -40dB, frequency response 50Hz-20kHz), voiceprint recognition module, vibration unit (vibration frequency 120Hz, amplitude 0.6g), micro energy storage unit (200mAh), RNNoise noise reduction module, local voice recognition module; 2. Steps to be followed: S1. The dual-microphone array collects user voice and air conditioner noise signals in the living room at a sampling frequency of 16kHz and a sampling precision of 16bit. AGC normalizes the signal amplitude to 0.2V. The piezoelectric unit converts the air conditioner noise vibration into electrical energy (conversion efficiency of 18%), which is stored in the energy storage unit to power the microphone array in standby mode. Noise reduction is achieved by LMS algorithm, with a noise reduction of 12dB. S2. The noise sensor detects an ambient noise level of 45dB and sets the trigger threshold to 0.5 according to the mapping table; it calls the RNNoise model to filter low-frequency noise of 50Hz-150Hz generated by the air conditioner with a filtration efficiency of 90% and removes the silent segment by endpoint detection. S3. The local speech recognition model converts the speech signal into text, recognizes the keyword "song change", extracts it and matches it with the standard instruction library, reads the current scene label as "music", and generates a standardized "song change" control instruction; S4. Extract the 14-dimensional MFCC voiceprint features of the speech and compare them with the pre-stored user voiceprints. The similarity is 93%. Capture the user's lip movement image through the device's camera (15fps), extract 14 lip movement feature points, and successfully match them with the pre-stored lip shape library. The second verification is passed. Send the command to the speaker's audio unit to perform the song switching operation. S5. Synchronously outputs a short prompt tone (660Hz, 150ms) and two short vibrations (120Hz, 100ms each, 50ms interval) to allow the user to quickly confirm that the song has been switched. 3. Results data: Standby power consumption reduced by 40%, recognition accuracy in quiet scenes reached 95%, recognition accuracy in 70dB noise environment reached 90%, false execution rate was 3%, and user interaction satisfaction improved by 38%.

[0022] Comparative Example: Traditional Fixed Threshold Voice Control 1. Conditions: Single microphone (pickup sensitivity -40dB), fixed speech recognition threshold of 0.5, no adaptive noise reduction, no multimodal verification, single auditory feedback, no noise energy recovery, using a cloud-based speech recognition model (inference latency 1.5s). 2. Test environment: 70dB noise environment (simulating home air conditioner noise), the test command is "pause music" and "change song", and the test is repeated 100 times; 3. Results: The speech recognition accuracy was 72%, the false trigger rate was 28%, the operation error rate in screenless scenarios was 35%, there was no energy recovery function, the standby power consumption was not reduced, the response time was 1.8s, and the instruction confusion rate in multi-audio stream scenarios was 25%, which could not meet the usage requirements of high accuracy and low false trigger.

[0023] Effect Comparison Table Test metrics Embodiments of the present invention Traditional Comparison Increase 85dB noise recognition accuracy 88% 65% 23% Command false trigger rate 5%-8% 30%-35% ≥70% Operation accuracy in screenless scenarios 98% 65% 33% Reduced standby power consumption 40%-42% 0% 40%+ Response time 0.1s 1-2s 90%+ Multi-audio stream instruction obfuscation rate 0% 25% 100% In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0024] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. An audio stream control method based on voice interaction, characterized in that, Includes the following steps: S1. A microphone array consisting of 3-6 piezoelectric pickup units is used to collect user voice signals and environmental noise signals in real time. The signal amplitude is normalized to 0.1V-0.5V, and then preliminary noise reduction is performed through an adaptive interference cancellation algorithm to remove some fixed frequency noise. S2. The ambient noise decibel value is detected in real time by a noise sensor at a frequency of 1 time / 100ms. A corresponding mapping relationship between noise decibel and speech recognition trigger threshold is established. The trigger threshold is adaptively adjusted according to the real-time detected noise decibel. At the same time, the RNNoise deep learning noise reduction model is called to perform time-domain-frequency domain joint filtering on the preprocessed signal to remove non-human low-frequency noise (50Hz-200Hz) and retain effective speech signal (300Hz-3400Hz). S3. Input the noise-reduced speech signal into the local lightweight speech recognition model to complete the speech-to-text operation. Extract keywords and perform semantic analysis on the recognized text, match the preset standard instruction library, and read the scene tags of the audio stream currently running on the device. Combine the scene tags and semantic analysis results to accurately distinguish the instruction control object and generate standardized control instructions. S4. Extract voiceprint features from the identified speech signal. The extracted voiceprint features include 12-16 dimensional feature parameters such as fundamental frequency, spectral envelope, and Mel-frequency cepstral coefficients (MFCC). Compare these features with the user's voiceprint key pre-stored on the device. If the similarity is ≥90%, the user is deemed legitimate. Based on this, a secondary verification is performed using the user's biometrics. The secondary verification uses either laryngeal electromyography (EMG) signal verification or lip movement feature verification. After the verification is successful, standardized control commands are sent to the audio playback unit to execute the corresponding audio flow control operation. S5. After the command is executed, auditory sound feedback and vibration tactile feedback are output simultaneously. The auditory sound feedback uses emotional sound effects, and different sound effect parameters correspond to different control commands. The vibration tactile feedback is achieved through the device's built-in vibration unit, and different vibration modes correspond to different control commands, ensuring that users can quickly confirm the operation results in screenless scenarios.

2. The audio flow control method based on voice interaction according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S11. The piezoelectric pickup unit of the microphone array is made of lead zirconate titanate (PZT) piezoelectric material, with a pickup sensitivity of -38dB±2dB and a frequency response range of 50Hz-20kHz. While acquiring the user's voice signal, it simultaneously acquires the mechanical vibration signal generated by the ambient noise. S12. The collected noise vibration signal is converted into DC power through the piezoelectric effect with a conversion efficiency of 15%-25%. The converted power is stored in a micro energy storage unit (capacity of 100mAh-500mAh) for standby power supply of the microphone array. S13. The converted noise vibration electrical signal is filtered and amplified, with the amplification factor set to 20dB-40dB. The processed signal is analyzed by the adaptive least mean square (LMS) algorithm to generate an inverse filter signal with the same amplitude and opposite phase as the noise signal. S14. The reverse filtered signal is superimposed onto the original acquired mixed speech and noise signal to achieve active noise reduction, with a noise reduction of 10dB-18dB. At the same time, the superimposed signal is de-aliased to avoid signal distortion.

3. The audio stream control method based on voice interaction according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S21. The noise sensor uses an electret microphone sensor with a measurement range of 30dB-120dB and a measurement accuracy of ±1dB. It detects the ambient noise decibel value in real time and updates the detection data every 100ms. S22. Establish a decibel-threshold mapping table, with the following specific mapping relationships: When the noise level is 0-40dB (quiet scene), the speech recognition trigger threshold is set to 0.3 (normalized amplitude); when the noise level is 40-60dB (everyday scene), the trigger threshold is set to 0.5 (normalized amplitude); when the noise level is 60-85dB (noisy scene), the trigger threshold is set to 0.7 (normalized amplitude); when the noise level is above 85dB (strong noise scene), the trigger threshold is set to 0.8 (normalized amplitude), and the ultrasonic active anti-interference mechanism is activated simultaneously. S23. Call the RNNoise deep learning model, which adopts a recurrent neural network (RNN) architecture and includes an input layer, hidden layers (3-5 layers), and an output layer. The input layer receives the preprocessed speech signal, the hidden layer extracts features through the activation function (ReLU), and the output layer outputs the denoised clean speech signal, with a filtering efficiency of ≥90% for non-human voice low-frequency noise in the 50Hz-200Hz range. S24. Perform endpoint detection on the denoised speech signal, and use a combination of short-time energy and short-time zero-crossing rate to accurately distinguish between speech segments and silence segments, remove silence segment signals, reduce invalid data processing, and improve system response speed.

4. The audio flow control method based on voice interaction according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S31. The local lightweight speech recognition model adopts an architecture that combines deep neural networks (DNN) and hidden Markov models (HMM). The model training dataset contains more than 100,000 speech command data with different accents and different noise environments, with a recognition accuracy of ≥95% (quiet scene). S32. Extract keywords from the speech-to-text recognition results. Use the TF-IDF algorithm to extract keywords with an extraction threshold of 0.

6. Match the extracted keywords with the standard instruction library. If the similarity is ≥85%, it is considered a valid instruction. S33. The device records the currently running audio stream information in real time and assigns a unique scene tag to each audio stream. Scene tags include music, video, podcast, and audiobook. The tag information is stored in the scene management module on the device and can be updated in real time. S34. Combining keyword matching results with scene tags, perform semantic intent analysis. When the keyword "pause" is identified and the scene tag is "music", generate a standardized control command to "pause music"; when the scene tag is "video", generate a standardized control command to "pause video" to avoid confusion in multi-audio stream control.

5. The audio flow control method based on voice interaction according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S41. Voiceprint feature extraction adopts the Mel frequency cepstral coefficient (MFCC) method to extract 16-dimensional feature parameters. The sampling window length is set to 25ms and the frame shift is set to 10ms. The extracted feature parameters are normalized to remove the influence of individual differences. S42. Voiceprint comparison uses the Dynamic Time Warping (DTW) algorithm, with a comparison threshold of 0.

1. When the similarity is ≥90%, the user is considered legitimate; if the similarity is <90%, the control command is refused and a prompt sound effect is output. S43. When the secondary verification uses laryngeal electromyography (EMG) signals, the device's built-in EMG sensor collects the micro-vibration signals of the laryngeal muscles when the user speaks. The collection frequency is 1kHz. The peak value and effective value features of the EMG signals are extracted and compared with the pre-stored EMG feature templates. If the similarity is ≥88%, the verification passes. S44. When the secondary verification uses lip movement feature verification, the device camera captures the lip movement image of the user speaking in real time at a frame rate of 15fps. The lip movement image is preprocessed (grayscale, noise reduction, normalization), and lip movement feature points (12-16) are extracted and matched with the pre-stored lip movement mouth shape library. If the match is successful, the verification is passed. S45. Only when both voiceprint verification and secondary verification pass will the standardized control command be sent to the audio playback unit to execute audio flow control operation. If the verification fails, the operation will be terminated to ensure control security.

6. The audio stream control method based on voice interaction according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: S51. The specific parameters for auditory sound effect feedback are as follows: the play command corresponds to an upward pitch, with the frequency rising from 440Hz to 880Hz and a duration of 300ms; the pause command corresponds to a fade-out sound effect, with the frequency falling from 880Hz to 440Hz and a duration of 300ms; the skip command corresponds to a short prompt tone, with a frequency of 660Hz and a duration of 150ms; the volume adjustment command corresponds to a continuous sound effect, with a frequency of 550Hz and a duration synchronized with the adjustment duration. S52. Vibration tactile feedback is achieved through an eccentric motor vibration unit, with a vibration frequency set to 100Hz-200Hz and a vibration amplitude set to 0.5g-1g. Specifically, the pause command corresponds to one short vibration with a duration of 100ms; the song skip command corresponds to two short vibrations, each with a duration of 100ms and an interval of 50ms; and the volume adjustment command corresponds to a long vibration with a duration consistent with the adjustment operation duration (not exceeding 2s). S53. Sound and vibration feedback are activated simultaneously with a startup delay of ≤50ms, ensuring that users can receive both auditory and tactile feedback at the same time, thus improving the accuracy of operation confirmation in screenless scenarios.

7. The audio stream control method based on voice interaction according to claim 2, characterized in that, The micro energy storage unit uses a lithium polymer battery with a charge-discharge cycle life of ≥1000 times and an operating temperature range of -10℃ to 60℃. When the energy storage unit's charge is ≥80%, noise and vibration energy recovery is stopped; when the charge is ≤20%, the energy recovery priority enhancement mode is activated, and the energy recovery efficiency is increased to 30% to maintain the microphone array's standby power supply, which can reduce the device's standby power consumption by 40%-42%.

8. The audio stream control method based on voice interaction according to claim 3, characterized in that, In a high-noise environment of 85dB, after dynamic threshold adjustment and RNNoise model noise reduction, the speech recognition accuracy is ≥88%, and the false trigger rate is reduced by ≥30% compared with the traditional fixed threshold scheme. In a strong noise environment of 100dB, the ultrasonic active anti-interference mechanism is activated to emit 25-40kHz ultrasonic waves to form an acoustic barrier, which can maintain the speech recognition accuracy above 75%.

9. The audio flow control method based on voice interaction according to claim 5, characterized in that, After dual modal verification using both voiceprint and biometric features, the command error rate is reduced by ≥90%, effectively preventing others from impersonating voice commands for control. It is suitable for high-security control scenarios such as vehicle payment and smart home access control, meeting financial-grade security requirements.

10. The audio flow control method based on voice interaction according to claim 6, characterized in that, By combining auditory sound effects with tactile vibration feedback, the accuracy of operation confirmation in screenless scenarios (driving, sports) is improved by ≥50%. Compared with single auditory feedback, user interaction satisfaction is improved by ≥35%. Psychological experiments have verified that the user operation error rate is reduced by more than 60%.

Images