Voice recognition method and system of AI intelligent voice remote controller
By performing frequency and time domain analysis on audio signals, combining fundamental frequency components and tailing effects, and using a neural network model for speech recognition, the problem of insufficient accuracy of traditional massage machine speech recognition in complex environments is solved, achieving a highly efficient speech recognition effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-03-27
AI Technical Summary
Traditional massage machines' voice recognition technology is easily affected by external noise and multiple sound sources in complex environments, leading to a decrease in voice recognition accuracy. This results in the inability to accurately recognize user interaction voice commands, increasing device energy consumption and reducing user experience.
By performing comprehensive analysis of the audio signal in both the frequency and time domains, the short-time energy correlation strength is evaluated. Combining the autocorrelation strength and tailing effect of the fundamental frequency component, a neural network model is used for speech recognition to extract the human voice signal.
It improves the accuracy and robustness of speech recognition in complex environments, enhances the ability to recognize low sound pressure and weak speech, optimizes the accuracy and response speed of speech recognition, and improves the user interaction experience.
Smart Images
Figure CN121747564A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of speech recognition technology, specifically to a speech recognition method and system for an AI smart voice remote control. Background Technology
[0002] With the rapid development of AI and IoT technologies, massage machines, as an important product in the health and wellness industry, are gradually evolving from functional to intelligent, personalized, and convenient features. Traditional massage machines rely heavily on physical remote controls, buttons on the machine itself, or mobile apps for operation. Users often need to frequently press buttons on the remote or go through complex procedures to switch between and adjust different massage modes. The introduction of AI voice interaction technology not only frees users' hands completely but also enables personalized, precise, and rapid massage adjustments.
[0003] Currently, in the integration of smart home appliances with smart voice control, the multiple sound sources generated by the operation of the massager itself and the reverberation in the enclosed space threaten the accuracy of voice recognition. Traditional algorithms for acquiring and recognizing human voice interaction audio signals are easily affected by external environmental noise and multiple sound sources generated by the operation of the massager itself. Relying solely on short-time energy and zero-crossing rate cannot determine the threshold of active frames for human voice audio, thus failing to accurately identify the start and end points of speech. This makes it difficult to effectively extract the sound sources of human voice interaction, significantly reducing the recognizability of human voice audio. As a result, the voice remote control may be repeatedly and falsely activated or may fail to fully recognize user interaction voice commands, increasing device energy consumption and reducing user experience. Summary of the Invention
[0004] To address the aforementioned technical problems, the purpose of this application is to provide a voice recognition method and system for an AI smart voice remote control, the specific technical solution of which is as follows: In a first aspect, embodiments of this application provide a voice recognition method for an AI smart voice remote control, the method comprising the following steps: Audio signals are collected in real time via a voice remote control, and frequency domain transformation is performed on each frame of audio signal. Analyze the energy fluctuation and energy shift of each frame of audio signal in the frequency domain, the frequency of zero crossing of the amplitude of each frame of audio signal in the time domain, and the autocorrelation strength of the energy within the fundamental frequency component of each frame of audio signal in the frequency domain to determine the short-time energy correlation strength of each frame of audio signal. Identify the tail envelopes within the fundamental frequency components of each frame of audio signal, measure the fundamental frequency tail effect of the audio signal by the change trend of the energy corresponding to the harmonic frequencies of each frame of audio signal in the frequency domain, and the width of each tail envelope, and obtain the pulse harmonic attenuation of each frame of audio signal by combining the peak shape distribution of each frame of audio signal in the time domain. By combining the short-time energy correlation intensity and the pulse harmonic attenuation, the saliency of human voice in each frame of audio signal is constructed, thereby extracting the human voice signal from the collected audio signal and performing speech recognition on the human voice signal.
[0005] In one embodiment, determining the short-time energy correlation strength of each frame of audio signal includes: Determine the degree of energy dispersion of all frequency components of each frame of audio signal in the frequency domain, and count the number of times the amplitude of each frame of audio signal crosses zero in the time domain; The autocorrelation of the sequence of energy components of all frequency components within the fundamental frequency component of each frame of audio signal is analyzed, and the skewness of the sequence of energy components of all frequency components of each frame of audio signal in the frequency domain is calculated. The short-time energy correlation strength of each frame of audio signal is positively correlated with the degree of dispersion, the number of times, and the autocorrelation, and negatively correlated with the skewness.
[0006] In one embodiment, the autocorrelation is measured by the mean of all autocorrelation coefficients of the sequence of energy of all frequency components within the fundamental frequency component of each frame of audio signal in the frequency domain.
[0007] In one embodiment, the short-time energy correlation intensity is calculated as follows: In the formula, Let be the short-time energy correlation strength of the audio signal in the i-th frame. Let represent the degree of energy dispersion of all frequency components of the i-th frame audio signal in the frequency domain. Let be the number of times the i-th frame of the audio signal crosses zero in the time domain amplitude. Let be the mean of all autocorrelation coefficients of the sequence composed of the energy of all frequency components in the fundamental frequency component of the i-th frame audio signal in the frequency domain. Let be the skewness of the sequence of energy components of all frequency components of the i-th frame audio signal in the frequency domain, and norm() be the normalization function.
[0008] In one embodiment, identifying each trailing envelope within the fundamental frequency component of each frame of audio signal includes: The peak lookup algorithm is used to obtain the energy peaks of each frame of audio signal in the fundamental frequency component in the frequency domain. The signal envelopes corresponding to all remaining energy peaks except the maximum energy peak are used as the tail envelopes.
[0009] In one embodiment, the measurement of the fundamental frequency tail effect of the audio signal includes: Obtain the 3dB bandwidth of each trailing envelope, perform linear fitting on the energy corresponding to all harmonic frequencies of each frame audio signal in the frequency domain, and obtain the slope of the fitted line; calculate the product of the absolute value of the slope and the number of all harmonic frequencies of each frame audio signal in the frequency domain, and perform inverse proportional mapping on the product result. The results of the inverse proportional mapping, the number of trailing envelopes corresponding to each frame of audio signal, and the 3dB bandwidth of the trailing envelope are positively fused to obtain the trailing harmonic attenuation factor of each frame of audio signal, which is used to measure the fundamental frequency trailing effect of the audio signal.
[0010] In one embodiment, obtaining the pulse harmonic attenuation of each frame of audio signal includes: Obtain all peaks of each frame of audio signal in the time domain, and take the time interval between each peak and the nearest zero point before it as the pulse time interval of each peak; calculate the ratio of the amplitude of each peak to its pulse time interval, and take the sum of the ratios of all peaks of each frame of audio signal in the time domain as the time domain narrowband pulse intensity of each frame of audio signal. The normalized value of the product of the trailing harmonic attenuation factor of each frame of audio signal and the time-domain narrowband pulse intensity is used as the pulse harmonic attenuation degree of each frame of audio signal.
[0011] In one embodiment, the salience of human voice in each frame of audio signal is the ratio of the short-time energy correlation intensity to the pulse harmonic attenuation of each frame of audio signal.
[0012] In one embodiment, the speech recognition of the human voice signal includes: The saliency of human voices in all collected audio frames is thresholded, and audio frames with saliency greater than or equal to the threshold are used as human voice signals. A neural network model is then used to perform speech recognition on the human voice signals.
[0013] Secondly, embodiments of this application also provide a voice recognition system for an AI smart voice remote control, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.
[0014] This application has at least the following beneficial effects: This application performs comprehensive frequency and time domain analysis on each frame of audio signal to evaluate short-time energy correlation strength, accurately reflecting the energy changes of human voice in the audio signal. By analyzing the energy fluctuations and shifts in the frequency domain, it solves the problem of traditional speech recognition's inability to effectively distinguish weak speech signals under low energy and background noise, enhancing the recognition ability for low sound pressure and weak speech in complex environments and effectively improving the recognition rate in noisy environments. Furthermore, by extracting the fundamental frequency components of the audio signal and analyzing the autocorrelation strength, it can effectively distinguish the signal features of human voice from non-human voice, helping to avoid misjudging noise or non-human voice components as valid speech, especially in cases of complex background noise or poor sound quality. Utilizing the changing trends of harmonic frequencies in the frequency domain, it can more accurately capture the basic features of speech, improving the robustness of speech recognition. By analyzing the fundamental frequency trailing effect of each frame of audio signal, the width of the trailing envelope and its impact on the audio signal can be accurately measured and evaluated. By measuring the changing trend of the trailing effect, it helps to avoid confusion or misrecognition caused by the trailing effect in the speech signal. Especially in the case of fast speech or continuous pronunciation, the trailing effect often affects the clarity and accuracy of recognition. This application significantly improves the ability to capture details in the speech signal and optimizes the accuracy of speech recognition. The measurement of pulse harmonic attenuation, combined with short-time energy analysis, can accurately capture the frequency dynamics of speech, which helps to improve the response speed to speech dynamics. It solves the problem of the impact of background noise interference on the speech recognition effect in complex environments, effectively extracts the human voice interaction sound source from the audio signal, improves the recognizability of human voice audio, and ensures the user's voice interaction experience. Attached Figure Description
[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart illustrating the steps of a voice recognition method for an AI smart voice remote control, as provided in one embodiment of this application. Detailed Implementation
[0017] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a voice recognition method and system for an AI smart voice remote control proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0019] The following description, in conjunction with the accompanying drawings, details the specific solution of the voice recognition method and system for an AI intelligent voice remote control provided in this application.
[0020] Please see Figure 1 The diagram illustrates a flowchart of a voice recognition method for an AI smart voice remote control according to an embodiment of this application. The method includes the following steps: S1 can collect audio signals in real time via a voice remote control.
[0021] In this embodiment, a supercardioid MEMS microphone is deployed at the front end of the voice remote control of the massager. The microphone audio signal is acquired via an I2S / TDM interface at a sampling frequency of 16kHz and a precision of 16bit. The acquired microphone audio signal is processed into frames of 20ms each (320 samples), with the frame shift set to half the frame length. A Hamming window is selected to window the microphone audio signal in each frame. A bandpass filter is used to filter out high-frequency noise exceeding half the oversampling frequency to avoid signal aliasing. The acquired microphone audio signal is converted into a digital signal via an ADC analog-to-digital converter. Since framing and windowing are well-known technologies, the specific acquisition process will not be described in detail.
[0022] S2, analyze the degree of energy fluctuation and energy shift of each frame of audio signal in the frequency domain, as well as the frequency of zero crossing of the amplitude of each frame of audio signal in the time domain, and the autocorrelation strength of the energy within the fundamental frequency component of each frame of audio signal in the frequency domain, to determine the short-time energy correlation strength of each frame of audio signal.
[0023] During the operation of a massage machine, the microphone audio signal acquired by the smart voice remote control typically contains multiple sound source signals, including human voice audio signals, internal motor vibration audio signals, mechanical friction sound audio signals, and sound reflection audio signals from enclosed spaces. Traditional VAD endpoint detection algorithms rely solely on short-time energy and zero-crossing rate to determine the threshold of active human voice audio frames. This makes it impossible to accurately extract active frames of human voice signals from continuous multiple sound source signals, which may result in inaccurate extraction of the start and end points of active frames of human voice signals. Consequently, the voice remote control's ability to recognize human voice interaction sound sources decreases, reducing the user experience.
[0024] Specifically, during the operation of the massage machine, when the microphone audio signal in any frame acquired by the voice remote control is more likely to contain human voice interaction source signals, the short-time energy entropy of the microphone audio signal in the signal frame is stronger due to the syllable discontinuity of the human voice interaction audio signal and the stability of the massage machine operation audio signal. The phenomenon of the signal frequently crossing the zero point is more obvious. At the same time, the energy distribution of the frequency components of the human voice signal is more concentrated in the mid-low frequency, while the frequency components of the massage machine motor and mechanical friction audio signal are more concentrated in the high frequency. Therefore, the stronger the autocorrelation of the fundamental frequency of the microphone audio signal in the signal frame, the more obvious the shift of the spectral energy center to the low frequency region.
[0025] Based on the above analysis, this embodiment calculates the short-time energy correlation intensity of each frame of audio signal to characterize the high-frequency state of short-time energy and the degree of spectral period shift in each frame of microphone audio signal acquired by the voice remote control. Specifically: First, the time-domain audio signal is analyzed, and the number of zero-crossing points of the signal amplitude in each frame of the microphone audio signal is counted. Second, the amplitude time sequence of each frame of the microphone audio signal acquired by the voice remote control is used as input, and the short-time four-dimensional transform (FFT) is used to obtain the frequency domain representation of each frame of the microphone audio signal. Then, the Hilbert transform is used to obtain all frequency components of each frame of the microphone audio signal in the frequency domain, where the fundamental frequency component contains the highest value of the frequency domain response energy. Both the short-time four-dimensional transform (FFT) and the Hilbert transform are existing well-known techniques, and the specific process will not be elaborated upon.
[0026] For each frame of microphone audio signal acquired by the voice remote control, the energy of all frequency components within its fundamental frequency component is arranged in ascending order and denoted as the fundamental frequency energy sequence. All autocorrelation coefficients of the fundamental frequency energy sequence are calculated, and the mean of these autocorrelation coefficients is denoted as the energy period intensity of the fundamental frequency component, reflecting the autocorrelation of the fundamental frequency energy sequence. Furthermore, the energy of all frequency components of each frame of microphone audio signal acquired by the voice remote control in the frequency domain is arranged in ascending order and denoted as the frequency energy sequence of each frame of microphone audio signal.
[0027] The short-time energy correlation strength of each frame of audio signal is calculated using the following expression: In the formula, Let be the short-time energy correlation strength of the audio signal in the i-th frame. The dispersion of the energy of all frequency components of the i-th frame audio signal in the frequency domain is used to measure the energy fluctuation. In this embodiment, the dispersion is calculated using information entropy. Implementers can choose other existing methods, such as variance, standard deviation, coefficient of variation, etc. Let be the number of times the i-th frame of the audio signal crosses zero in the time domain amplitude. Let be the mean of all autocorrelation coefficients of the sequence of energy components of the fundamental frequency component of the i-th frame audio signal in the frequency domain, i.e., the energy periodic intensity of the fundamental frequency component of the i-th frame audio signal, where norm() is the normalization function. The skewness is the sequence of energy components of all frequency components in the i-th frame of the audio signal in the frequency domain. A lower skewness indicates a stronger leftward bias in the frequency energy sequence, which manifests as a more pronounced shift of the response energy of all frequency components towards lower frequencies in the microphone audio signal frequency domain. It should be noted that the norm normalization function in this embodiment uses the maximum-minimum normalization method, that is, normalizing the short-time energy correlation intensity of all frames of audio signals. When the skewness... When the frequency energy sequence is perfectly symmetrically distributed, and the influence of the audio signal's frequency domain offset is ignored, the short-time energy correlation strength of the i-th frame audio signal is... The calculation method is as follows .
[0028] It should be understood that the short-time energy correlation strength reflects the high-frequency characteristics of short-time energy and the degree of spectral period shift in each frame of microphone audio signal acquired by the voice remote control; dispersion. Number of zero crossings This reflects the energy fluctuation of the microphone audio signal acquired by the voice remote control in the frequency domain across all frequency components, as well as the frequency of zero-crossing points of the microphone audio signal amplitude; while the mean autocorrelation coefficient... This characterizes the autocorrelation strength of the fundamental frequency component in each frame of microphone audio signal acquired by the voice remote control. Skewness reflects the high-frequency shift of the frequency component energy. During the operation of the massage machine, when there is a greater likelihood of human voice interaction source signal in any frame of microphone audio signal acquired by the voice remote control, the short-time energy entropy of the microphone audio signal caused by the discontinuity of the human voice interaction audio signal syllables is greater, the frequency of the microphone audio signal amplitude crossing zero is stronger, and at the same time, the periodic autocorrelation strength of the fundamental frequency of the microphone audio signal in the signal frame is greater, and the shift of the spectral energy center to the low-frequency region is more obvious, that is, the greater the short-time energy correlation strength of the final audio signal.
[0029] S3 identifies each trailing envelope within the fundamental frequency component of each frame of audio signal. By measuring the change trend of the energy corresponding to the harmonic frequency in the frequency domain of each frame of audio signal, as well as the width of each trailing envelope, the fundamental frequency trailing effect of the audio signal is measured. Combined with the peak shape distribution of each frame of audio signal in the time domain, the pulse harmonic attenuation of each frame of audio signal is obtained.
[0030] In the process of voice recognition in the AI smart voice remote control of the massage machine, it is also necessary to consider the interference between transient sound sources generated by the switching of the massage machine's operating mode and human voice interaction audio, as well as the impact of sound reflection audio signals in a closed space on the misjudgment of the start and end points of the voice. Multi-dimensional characteristics of harmonic structure and reflection attenuation should be analyzed to distinguish between transient mechanical structure noise and closed reverberation tails and real human voice interaction audio. Otherwise, it may still cause serious drift in the recognition of the start and end point of the voice and false triggering of the voice remote control, which will increase the energy consumption of the device and reduce the user experience.
[0031] Specifically, during the operation of the massage machine, when any frame of microphone audio signal acquired by the voice remote control is more likely to be mechanical transient impact noise or closed reverberation tail rather than a real human voice interaction audio signal, the microphone audio signal in the signal frame will exhibit a narrow and high pulse characteristic in the time domain due to the mechanical transient impact audio signal caused by the switching or shifting of the massage machine's operating mode. That is, the high and narrow peaks of the microphone audio signal will be more obvious. At the same time, the closed reverberation tail will cause the audio signal energy to exhibit a trailing phenomenon. The fundamental frequency trailing effect of the microphone audio signal in the signal frame is severe, and the harmonic components in the microphone audio signal in the signal frame are more due to the slow release of energy by the closed reverberation tail, and the harmonic frequency components decay more slowly.
[0032] Based on the above analysis, this embodiment first uses the temporal amplitude sequence of each frame of microphone audio signal acquired by the voice remote control as input. It then employs the AMPD (Automatic Multiscale-based Peak Detection) algorithm to acquire all peak signals. For any given peak signal, it acquires the first zero point closest to that peak signal before its corresponding moment, denoted as the starting point of that peak signal. The zero point is the moment when the amplitude value is 0. The time interval between each peak signal and its corresponding starting point is calculated and denoted as the pulse time interval of each peak signal. The smaller the pulse time interval, the more pronounced the time-domain narrowband effect when the peak signal reaches its amplitude peak.
[0033] Therefore, in this embodiment, the ratio of the amplitude of each peak signal in the time domain to its pulse time interval is calculated for each frame of audio signal. The sum of the ratios of all peaks in the time domain of each frame of audio signal is used as the time-domain narrowband pulse intensity of each frame of audio signal. The greater the time-domain narrowband pulse intensity, the more obvious the time-domain narrowband effect is when the peak signal reaches the peak amplitude.
[0034] Secondly, the AMPD multi-scale peak algorithm is used to obtain all energy peaks of the fundamental frequency component corresponding to each frame of the microphone audio signal. The signal envelopes corresponding to all energy peaks in the fundamental frequency component, except for the largest energy peak, are denoted as the tail envelopes of the fundamental frequency component. The 3dB bandwidth of all tail envelopes in the fundamental frequency component is calculated. All integer multiples of the fundamental frequency in each frame of the microphone audio signal are obtained and denoted as harmonic frequencies of each frame of the microphone audio signal. When the energy of the Nth integer multiple of the fundamental frequency is 0, it indicates that there are no harmonic frequencies greater than or equal to the Nth integer multiple, where N represents the multiple of the fundamental frequency. The sequence of energies corresponding to all harmonic frequencies is used as input, and the least squares method is used to obtain the fitted line of the energy corresponding to the harmonic frequencies. In the fitting process, the horizontal axis is the harmonic frequency, and the vertical axis is the energy. The slope of the obtained fitted line is calculated. The larger the slope, the slower the energy decay of the harmonic frequency.
[0035] Furthermore, the number of all harmonic frequencies corresponding to the fundamental frequency of each frame of audio signal is counted. The absolute value of the slope of the fitted straight line corresponding to each frame of audio signal is multiplied by the number of harmonic frequencies. The multiplication result is then subjected to an inverse proportional mapping. This inverse proportional mapping result is then positively fused with the number of trailing envelopes and the 3dB bandwidth of the trailing envelopes corresponding to each frame of audio signal to obtain the trailing harmonic attenuation factor of each frame of audio signal, which is used to measure the fundamental frequency trailing effect of the audio signal. It should be noted that the inverse proportional mapping means that the multiplication result and the trailing harmonic attenuation factor are negatively correlated, which can be calculated by using the reciprocal, the negative exponent of the natural constant e, etc.; positive fusion means combining multiple variables, which can be calculated by adding, multiplying, adding and multiplying together, or taking the mean.
[0036] In this embodiment, for each frame of audio signal, the average 3dB bandwidth of all corresponding trailing envelopes is calculated. The reciprocal of the multiplication result is then multiplied by the number of trailing envelopes and the average value, which serves as the trailing harmonic attenuation factor for each frame of audio signal. It should be noted that when the number of harmonic frequencies corresponding to the fundamental frequency of a certain frame of audio signal is less than or equal to one, it indicates that the degree of harmonic interference to that frame of audio signal is low. In this case, this embodiment uses the product of the number of trailing envelopes corresponding to that frame of audio signal and the average value as the trailing harmonic attenuation factor for that frame of audio signal.
[0037] Based on the above analysis, the pulse harmonic attenuation of each frame of audio signal is calculated to characterize the temporal transient pulse intensity and trailing harmonic attenuation in each frame of microphone audio signal acquired by the voice remote control. The specific expression is as follows: In the formula, The pulse harmonic attenuation of the i-th frame microphone audio signal acquired by the voice remote control; The time-domain narrowband pulse intensity of the microphone audio signal in the i-th frame acquired by the voice remote control. The trailing harmonic attenuation factor of the i-th frame microphone audio signal acquired by the voice remote control.
[0038] It should be understood that the time-domain narrowband pulse intensity reflects the narrow and high amplitude of the time-domain pulse in each frame of microphone audio signal acquired by the voice remote control; while the tailing harmonic attenuation factor characterizes the severity of the fundamental frequency tailing effect in each frame of microphone audio signal acquired by the voice remote control, as well as the degree of harmonic presence and the slow attenuation of harmonic frequencies. During the operation of the massage machine, when any frame of microphone audio signal acquired by the voice remote control is more likely to be mechanical transient impact noise or closed reverberation tail, rather than a real human voice interaction audio signal, the narrowband peak of the time-domain amplitude pulse in the signal frame is more obvious, that is, the calculated time-domain narrowband pulse intensity becomes larger; at the same time, due to severe interference from closed reverberation tail, the more harmonic frequency components there are in the microphone audio signal in the signal frame, the slower the attenuation of the harmonic frequency component response energy, that is, the calculated tailing harmonic attenuation factor becomes larger.
[0039] S4. Combining the short-time energy correlation intensity and the pulse harmonic attenuation, the saliency of human voice in each frame of audio signal is constructed, thereby extracting the human voice signal from the collected audio signal and performing speech recognition on the human voice signal.
[0040] During the operation of the massage machine, when the short-time energy correlation strength of the microphone audio signal in each frame of the microphone audio signal acquired by the voice remote control is higher and the pulse harmonic attenuation is lower, the microphone audio signal in the signal frame is more likely to contain human voice interaction sound source signal, and is not caused by mechanical transient impact or closed reverberation tail. In this case, the signal frame should be used as a frame in the human voice interaction sound source recognition section of the voice remote control for accurate recognition of user voice commands.
[0041] Therefore, this embodiment constructs a human voice saliency to evaluate the possibility of real human voice interaction audio in each frame of microphone audio signal acquired by the voice remote control. It can be obtained by short-time energy correlation intensity and pulse harmonic attenuation. Specifically, in this embodiment, the ratio of short-time energy correlation intensity to pulse harmonic attenuation of each frame of microphone audio signal is used as the human voice saliency of each frame of microphone audio signal. When the human voice saliency is larger, it indicates that there is more likely to be real human voice interaction speech in the signal frame, and it is not caused by mechanical transient impact or closed reverberation tail interference of the massage machine.
[0042] Furthermore, after the user utters the voice wake-up word of the massager, all frame microphone audio signals and their corresponding human voice salience are obtained through the above steps. The human voice salience of all frame microphone audio signals obtained by the voice remote control is used as input, and the Otsu threshold is used to obtain the segmentation threshold. All signal frames greater than or equal to the segmentation threshold are used as human voice signals and recorded as human voice interaction frames. The PSOLA (Pitch synchronous Overlap Add) pitch synchronous overlay algorithm is used to obtain the interactive voice segment spliced from all adjacent human voice interaction frames.
[0043] Each interactive speech segment is used as input, and a Deep Neural Network (DNN) is used to obtain the interactive command recognition results for each interactive speech segment. Adam is used as the optimization algorithm, and the Mean Absolute Error (MAE) is used as the loss function. Since both the PSOLA pitch synchronization superposition algorithm and the DNN are well-known technologies, the specific acquisition process will not be described in detail. Implementers can choose other existing feasible algorithms and neural network models for acquiring interactive speech segments. This embodiment does not impose any restrictions on this.
[0044] Based on the same inventive concept as the above method, this application embodiment also provides a voice recognition system for an AI smart voice remote control, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described voice recognition methods for an AI smart voice remote control.
[0045] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0046] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0047] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.
Claims
1. A voice recognition method for an AI smart voice remote control, characterized in that, The method includes the following steps: Audio signals are collected in real time via a voice remote control, and frequency domain transformation is performed on each frame of audio signal. Analyze the energy fluctuation and energy shift of each frame of audio signal in the frequency domain, the frequency of zero crossing of the amplitude of each frame of audio signal in the time domain, and the autocorrelation strength of the energy within the fundamental frequency component of each frame of audio signal in the frequency domain to determine the short-time energy correlation strength of each frame of audio signal. Identify the tail envelopes within the fundamental frequency components of each frame of audio signal, measure the fundamental frequency tail effect of the audio signal by the change trend of the energy corresponding to the harmonic frequencies of each frame of audio signal in the frequency domain, and the width of each tail envelope, and obtain the pulse harmonic attenuation of each frame of audio signal by combining the peak shape distribution of each frame of audio signal in the time domain. By combining the short-time energy correlation intensity and the pulse harmonic attenuation, the saliency of human voice in each frame of audio signal is constructed, thereby extracting the human voice signal from the collected audio signal and performing speech recognition on the human voice signal.
2. The voice recognition method for an AI smart voice remote control as described in claim 1, characterized in that, Determining the short-time energy correlation strength of each frame of audio signal includes: Determine the degree of energy dispersion of all frequency components of each frame of audio signal in the frequency domain, and count the number of times the amplitude of each frame of audio signal crosses zero in the time domain; The autocorrelation of the sequence of energy components of all frequency components within the fundamental frequency component of each frame of audio signal is analyzed, and the skewness of the sequence of energy components of all frequency components of each frame of audio signal in the frequency domain is calculated. The short-time energy correlation strength of each frame of audio signal is positively correlated with the degree of dispersion, the number of times, and the autocorrelation, and negatively correlated with the skewness.
3. The voice recognition method for an AI intelligent voice remote control as described in claim 2, characterized in that, The autocorrelation is measured by the mean of all autocorrelation coefficients of the sequence of energy of all frequency components within the fundamental frequency component of each frame of audio signal in the frequency domain.
4. The voice recognition method for an AI intelligent voice remote control as described in claim 3, characterized in that, The short-time energy correlation intensity is calculated as follows: In the formula, Let be the short-time energy correlation strength of the audio signal in the i-th frame. Let represent the degree of energy dispersion of all frequency components of the i-th frame audio signal in the frequency domain. Let be the number of times the i-th frame of the audio signal crosses zero in the time domain amplitude. Let be the mean of all autocorrelation coefficients of the sequence composed of the energy of all frequency components in the fundamental frequency component of the i-th frame audio signal in the frequency domain. Let be the skewness of the sequence of energy components of all frequency components of the i-th frame audio signal in the frequency domain, and norm() be the normalization function.
5. The voice recognition method for an AI intelligent voice remote control as described in claim 1, characterized in that, The identification of each trailing envelope within the fundamental frequency component of each frame of audio signal includes: The peak lookup algorithm is used to obtain the energy peaks of each frame of audio signal in the fundamental frequency component in the frequency domain. The signal envelopes corresponding to all remaining energy peaks except the maximum energy peak are used as the tail envelopes.
6. The voice recognition method for an AI intelligent voice remote control as described in claim 1, characterized in that, The measurement of the fundamental frequency trailing effect of the audio signal includes: Obtain the 3dB bandwidth of each trailing envelope, perform linear fitting on the energy corresponding to all harmonic frequencies of each frame audio signal in the frequency domain, and obtain the slope of the fitted line; calculate the product of the absolute value of the slope and the number of all harmonic frequencies of each frame audio signal in the frequency domain, and perform inverse proportional mapping on the product result. The results of the inverse proportional mapping, the number of trailing envelopes corresponding to each frame of audio signal, and the 3dB bandwidth of the trailing envelope are positively fused to obtain the trailing harmonic attenuation factor of each frame of audio signal, which is used to measure the fundamental frequency trailing effect of the audio signal.
7. The voice recognition method for an AI intelligent voice remote control as described in claim 6, characterized in that, The acquisition of the pulse harmonic attenuation of each frame of audio signal includes: Obtain all peaks of each frame of audio signal in the time domain, and take the time interval between each peak and the nearest zero point before it as the pulse time interval of each peak; calculate the ratio of the amplitude of each peak to its pulse time interval, and take the sum of the ratios of all peaks of each frame of audio signal in the time domain as the time domain narrowband pulse intensity of each frame of audio signal. The normalized value of the product of the trailing harmonic attenuation factor of each frame of audio signal and the time-domain narrowband pulse intensity is used as the pulse harmonic attenuation degree of each frame of audio signal.
8. The voice recognition method for an AI intelligent voice remote control as described in claim 1, characterized in that, The salience of human voice in each frame of audio signal is the ratio of the short-time energy correlation intensity to the pulse harmonic attenuation of each frame of audio signal.
9. The voice recognition method for an AI intelligent voice remote control as described in claim 1, characterized in that, The speech recognition of human voice signals includes: The saliency of human voices in all collected audio frames is thresholded, and audio frames with saliency greater than or equal to the threshold are used as human voice signals. A neural network model is then used to perform speech recognition on the human voice signals.
10. A voice recognition system for an AI intelligent voice remote control, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-9.