WO2021208287A1 - 用于情绪识别的语音端点检测方法、装置、电子设备及存储介质 - Google Patents

用于情绪识别的语音端点检测方法、装置、电子设备及存储介质 Download PDF

Info

Publication number
WO2021208287A1
WO2021208287A1 PCT/CN2020/104783 CN2020104783W WO2021208287A1 WO 2021208287 A1 WO2021208287 A1 WO 2021208287A1 CN 2020104783 W CN2020104783 W CN 2020104783W WO 2021208287 A1 WO2021208287 A1 WO 2021208287A1
Authority
WO
WIPO (PCT)
Prior art keywords
audio signal
human voice
network model
voice
feature
Prior art date
Application number
PCT/CN2020/104783
Other languages
English (en)
French (fr)
Inventor
王德勋
徐国强
Original Assignee
深圳壹账通智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳壹账通智能科技有限公司 filed Critical 深圳壹账通智能科技有限公司
Publication of WO2021208287A1 publication Critical patent/WO2021208287A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/78—Detection of presence or absence of voice signals
    • G10L25/84—Detection of presence or absence of voice signals for discriminating voice from noise
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G10L25/24—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being the cepstrum
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
    • G10L25/63—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination for estimating an emotional state
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/78—Detection of presence or absence of voice signals
    • G10L25/87—Detection of discrete points within a voice signal

Definitions

  • This application relates to the technical field of speech semantics of artificial intelligence, and more specifically, to a speech endpoint detection method, device, electronic device, and storage medium for emotion recognition.
  • VAD Voice Endpoint Detection
  • VAD technology is also used in some speech coding such as G.729B and AMR-NB.
  • G.729B and AMR-NB Some speech coding
  • These methods also have a higher accuracy in a stable noise environment, but in a low signal-to-noise ratio environment and non-stationary environment The detection effect is not ideal.
  • the purpose of this application is to provide a voice endpoint detection method, device, electronic device, and storage medium for emotion recognition that can accurately recognize the location of the endpoint of human speech in a low signal-to-noise ratio environment and a non-stationary environment .
  • this application provides a voice endpoint detection method for emotion recognition, including:
  • processing operations on the audio signal including: adding pure noise segments and vocal noise segments in various scenarios to the audio signal and randomly setting a signal-to-noise ratio;
  • the above-mentioned MFCC features and their second-order differential features are input into the neural network model, and the high-dimensional information of the audio signal and the associated features of the audio signal before and after are extracted.
  • the high-dimensional information is a highly abstract phoneme feature learned from the sample by the neural network model,
  • the correlation feature represents the time series dynamic relationship of multiple adjacent frames of the audio signal;
  • the audio signal is divided into a human voice part and a non-human voice part.
  • this application also provides a voice endpoint detection device for emotion recognition, including:
  • Acquisition module to collect audio signals
  • a signal processing module for processing the audio signals collected by the acquisition module, the processing operations include: adding pure noise segments and vocal noise segments in various scenarios to the audio signal and randomly setting the signal-to-noise ratio;
  • the first feature extraction module extracts the MFCC feature and the second-order difference feature of the audio signal processed by the signal processing module;
  • the second feature extraction module inputs the MFCC features extracted by the first feature extraction module and their second-order differential features into the neural network model, and extracts the high-dimensional information of the audio signal and the associated features of the audio signal before and after the high-dimensional information is a neural network
  • the highly abstract phoneme features learned by the model from the samples, and the associated features represent the temporal dynamic relationship of adjacent multiple frames of the audio signal;
  • the endpoint recognition module inputs the high-dimensional information and associated features of the audio signal extracted by the second feature extraction module into the fully connected network model to obtain the detection result of each frame of the audio signal, and the detection result includes human voice and non-human voice;
  • the segmentation module divides the audio signal into the human voice part and the non-human voice part according to the detection result of the audio signal.
  • the present application also provides an electronic device, the electronic device includes a memory and a processor, the memory stores a voice endpoint detection program for emotion recognition, the voice endpoint for emotion recognition When the detection program is executed by the processor, the steps of the voice endpoint detection method for emotion recognition are realized.
  • the present application also provides a computer-readable storage medium that includes a voice endpoint detection program for emotion recognition, and the voice endpoint detection program for emotion recognition is When the processor is executed, the steps of the voice endpoint detection method for emotion recognition are realized.
  • the voice endpoint detection method, device, electronic device, and storage medium for emotion recognition described in this application add pure noise segments and vocal noise segments in various scenarios to audio signals and randomly set the signal-to-noise ratio, and combine the MFCC features and their
  • the second-order differential features are input to the neural network model, the high-dimensional information of the audio signal and the associated features of the audio signal are extracted, and the high-dimensional information and associated features of the audio signal are input into the fully connected network model to enhance the model’s robustness in complex and changeable environments. Robustness and generalization ability improve the recognition of noise data in low signal-to-noise ratio and non-stationary environment in traditional VAD.
  • FIG. 1 is a schematic diagram of an application environment of a preferred embodiment of a voice endpoint detection method for emotion recognition according to the present application
  • FIG. 2 is a schematic diagram of modules of a preferred embodiment of a voice endpoint detection program for emotion recognition in FIG. 1;
  • Fig. 3 is a flowchart of a preferred embodiment of a voice endpoint detection method for emotion recognition according to the present application.
  • This application provides a voice endpoint detection method for emotion recognition, which is applied to an electronic device.
  • FIG. 1 is a schematic diagram of an application environment of a preferred embodiment of a voice endpoint detection method for emotion recognition in this application.
  • the electronic device 1 may be a terminal client with computing functions, such as a server, a mobile phone, a tablet computer, a portable computer, a desktop computer, and the like.
  • the electronic device 1 includes a memory 11, a processor 12, a network interface 13 and a communication bus 14.
  • the memory 11 includes at least one type of readable storage medium.
  • the at least one type of readable storage medium may be a non-volatile storage medium such as flash memory, hard disk, multimedia card, card-type memory, and the like.
  • the readable storage medium may be an internal storage unit of the electronic device 1, such as a hard disk of the electronic device 1.
  • the readable storage medium may also be an external memory of the electronic device 1, such as a plug-in hard disk or a smart memory card (Smart Media Card, SMC) equipped on the electronic device 1. Secure Digital (SD) card, flash card (Flash Card), etc.
  • SD Secure Digital
  • flash card Flash Card
  • the readable storage medium of the memory 11 is generally used to store the voice endpoint detection program 10 for emotion recognition installed in the electronic device 1 and the like.
  • the memory 11 can also be used to temporarily store data that has been output or will be output.
  • the processor 12 may be a central processing unit (CPU), a microprocessor, or other data processing chip, and is used to run program codes or process data stored in the memory 11, for example, execute for emotions. Recognized voice endpoint detection program 10 and so on.
  • CPU central processing unit
  • microprocessor or other data processing chip
  • the network interface 13 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface), and is usually used to establish a communication connection between the electronic device 1 and other electronic clients.
  • a standard wired interface and a wireless interface such as a WI-FI interface
  • the communication bus 14 is used to realize the connection and communication between these components.
  • FIG. 1 only shows the electronic device 1 with the components 11-14, but it should be understood that it is not required to implement all the illustrated components, and more or fewer components may be implemented instead.
  • the electronic device 1 may also include a user interface, and the user interface may include an input unit such as a keyboard (Keyboard), a voice input device such as a microphone (microphone) and other clients with voice recognition functions, and a voice output device such as audio and earphones.
  • the user interface may also include a standard wired interface and a wireless interface.
  • the electronic device 1 may also include a display, and the display may also be called a display screen or a display unit.
  • it may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an organic light-emitting diode (OLED) touch device, and the like.
  • the display is used for displaying information processed in the electronic device 1 and for displaying a visualized user interface.
  • the electronic device 1 further includes a touch sensor.
  • the area provided by the touch sensor for the user to perform a touch operation is called a touch area.
  • the touch sensor described here may be a resistive touch sensor, a capacitive touch sensor, or the like.
  • the touch sensor includes not only a contact type touch sensor, but also a proximity type touch sensor and the like.
  • the touch sensor may be a single sensor, or may be, for example, a plurality of sensors arranged in an array.
  • the electronic device 1 may also include logic gate circuits, sensors, audio circuits, etc., which will not be repeated here.
  • the memory 11 as a computer storage medium may include an operating system and a voice endpoint detection program 10 for emotion recognition; the processor 12 executes the emotion recognition stored in the memory 11
  • the voice endpoint detection program 10 times implements the following steps:
  • processing operations on the audio signal including: adding pure noise segments and vocal noise segments in various scenarios to the audio signal and randomly setting a signal-to-noise ratio;
  • the MFCC feature is Mel-scale Frequency Cepstral Coefficients (MFCC for short) which are cepstral parameters extracted in the Mel-scale frequency domain , Mel scale describes the non-linear characteristics of the human ear frequency, and the second-order difference feature of the MFCC feature is the dynamic relationship between the three adjacent frames of the audio signal;
  • MFCC Mel-scale Frequency Cepstral Coefficients
  • the above-mentioned MFCC features and their second-order differential features are input into the neural network model, and the high-dimensional information of the audio signal and the associated features of the audio signal before and after the audio signal are extracted.
  • the high-dimensional information is a highly abstract phoneme feature learned by the neural network model from a large number of samples .
  • the phoneme feature refers to the part-of-speech feature, context information, and tone feature expressed according to the phoneme duration information, and the associated feature represents the temporal dynamic relationship of adjacent multiple frames of the audio signal;
  • the extracted audio signal is divided into a human voice part and a non-human voice part.
  • the voice endpoint detection program 10 for emotion recognition can also be divided into one or more modules, and the one or more modules are stored in the memory 11 and executed by the processor 12 to complete This application.
  • the module referred to in this application refers to a series of computer program instruction segments that can complete specific functions.
  • FIG. 2 it is a functional block diagram of a preferred embodiment of the voice endpoint detection program 10 for emotion recognition in FIG. 1.
  • the voice endpoint detection program 10 for emotion recognition can be divided into a collection module 110, a signal processing module 120, a first feature extraction module 130, a second feature extraction module 140, an endpoint recognition module 150, and a segmentation module 160, among which:
  • the collection module 110 collects audio signals
  • the signal processing module 120 performs processing operations on the audio signals collected by the acquisition module 110, and the processing operations include: adding pure noise segments and vocal noise segments in various scenarios to the audio signal and randomly setting a signal-to-noise ratio;
  • the first feature extraction module 130 extracts the MFCC feature and the second-order difference feature of the audio signal processed by the signal processing module 120;
  • the second feature extraction module 140 inputs the MFCC features and their second-order differential features extracted by the first feature extraction module 130 into the neural network model, and extracts the high-dimensional information of the audio signal and the before and after associated features of the audio signal;
  • the endpoint recognition module 150 inputs the high-dimensional information and associated features of the audio signal extracted by the second feature extraction module 140 into the fully connected network model to obtain the detection result of each frame of the audio signal.
  • the detection result includes human voice and non-human voice ;
  • the dividing module 160 divides the audio signal into a human voice part and a non-human voice part according to the detection result of the audio signal.
  • it further includes an emotion recognition module 170, which inputs the audio signal of the human voice part into the voice emotion detection model, and outputs the result of voice endpoint detection for emotion recognition.
  • an emotion recognition module 170 which inputs the audio signal of the human voice part into the voice emotion detection model, and outputs the result of voice endpoint detection for emotion recognition.
  • the aforementioned endpoint recognition module 150 performs a smooth operation on the output result of the fully connected network model, and when a frame that is a human voice or a non-human voice endpoint is detected, it detects the first frame before and after the frame. For the set number of frames of data, only when at least a second set number of detection results are consistent with the detection results of the one frame, the frame is finally determined to be human voice or non-human voice.
  • the first feature extraction module 130 includes:
  • Pre-emphasis unit pre-emphasis the audio signal
  • Framing and windowing unit to frame and window the pre-emphasized audio signal
  • Transformation unit which performs fast Fourier transform on the framed and windowed audio signal
  • the smoothing processing unit smoothes the frequency spectrum of the audio signal after the fast Fourier transform through a triangular band-pass filter, and eliminates the effect of harmonics to highlight the formant of the original voice;
  • the MFCC obtaining unit calculates the logarithmic energy output by the triangular band-pass filter, and obtains the MFCC through the discrete cosine transform;
  • the second-order difference feature obtaining unit obtains the dynamic relationship between three adjacent frames through the relationship between the previous first-order difference and the latter first-order difference of the MFCC, so as to obtain the second-order difference feature.
  • it further includes a training module to train the neural network model and the fully connected network model, including:
  • the noise adding unit obtains the human voice audio signal from the AISHELL data set, and adds a random combination of the pure noise segment and the human voice noise segment.
  • the AISHELL data set is an open source database of Beijing Hill Shell Technology Co., Ltd.;
  • the training set construction unit collects noise signals and human voice data in various scenarios and randomly sets the signal-to-noise ratio, thereby obtaining the training set of the human voice audio signals;
  • the feature extraction unit extracts the MFCC feature and its second-order difference feature of the audio signal in the training set
  • the training unit inputs the above-mentioned MFCC features and their second-order differential features into the neural network model, performs training, and extracts the high-dimensional information of the audio signal and the associated features before and after the audio signal.
  • the high-dimensional information is the neural network model learned from a large number of samples.
  • the highly abstract phoneme features of the audio signal are obtained through the second-order differential feature of the audio signal to obtain the before and after correlation characteristics of the audio signal; the high-dimensional information and the correlation characteristics of the audio signal are input into the fully connected network model to obtain the detection result of each frame of the audio signal;
  • the parameter update unit uses Focal loss as a loss function to iteratively update the parameters of the fully connected network model and the neural network model.
  • the aforementioned parameter update unit further modifies the weight of samples in the training set before each model parameter update, so as to reduce the weight of easy-to-classify samples and increase the weight of difficult-to-classify samples.
  • this application also provides a voice endpoint detection method for emotion recognition.
  • FIG. 3 it is a flowchart of a preferred embodiment of voice endpoint detection for emotion recognition in this application.
  • the method can be executed by a device, and the device can be implemented by software and/or hardware.
  • the voice endpoint detection method for emotion recognition includes:
  • Step S1 collecting audio signals
  • Step S2 processing the audio signal, the processing operation includes: adding the pure noise section and the vocal noise section to the audio signal in a variety of scenarios, and randomly setting the signal-to-noise ratio, the multiple scenarios including office noise sets, Hand-recorded laboratory noise collection, telephone collection noise collection and face-to-face audit data, etc., for example, set the signal-to-noise ratio in the range of 20dB-40dB, and obtain the ratio of audio signal to noise signal according to the set signal-to-noise ratio.
  • the noise signal is pure noise
  • the composition of the segment and the vocal noise segment, the power of the noise signal is determined, the mixing ratio of the pure noise segment and the vocal noise segment is adjusted, and it is stacked with the audio signal to obtain the mixing ratio closest to the human voice (without submerging the audio signal);
  • Step S3 Extract the MFCC feature and its second-order differential feature of the processed audio signal.
  • the method for extracting the MFCC feature and its second-order differential feature includes: pre-emphasizing the noise-added audio signal through a high-pass filter, and The pre-emphasized audio signal is framed and windowed, the framed and windowed audio signal is subjected to fast Fourier transform, the audio signal after the fast Fourier transform is smoothed through the Mel frequency filter bank, and each filter is calculated Group output logarithmic energy, through discrete cosine transform (DCT) to obtain MFCC (MFCC feature of audio signal), extract the difference spectrum of MFCC, and obtain its second-order difference feature;
  • DCT discrete cosine transform
  • Step S4 input the above-mentioned MFCC feature and its second-order differential feature into the neural network model, and extract the high-dimensional information of the audio signal and the associated features before and after the audio signal;
  • Step S5 Input the extracted high-dimensional information and associated features of the audio signal into the fully connected network model to obtain the detection result of each frame of the audio signal.
  • the detection result includes human voice and non-human voice, for example, the human voice is 1.
  • the non-human voice is 0, which turns the audio signal into a binary sequence;
  • Step S6 according to the detection result of the audio signal, divide the audio signal into the human voice part and the non-human voice part, for example, stack the frames with the detection result of 1 in the audio signal, and stack the frames with the detection result of 0 , To achieve the division of the human voice part and the non-human voice part.
  • step S6 it further includes: inputting the audio signal of the human voice part into the voice emotion detection model, and outputting the result of voice endpoint detection for emotion recognition, for example, dividing the voice emotion into happiness, interest, disgust, fear, and pain ( Sadness) and anger, etc.
  • the voice emotion detection model may be a classification model constructed based on a neural network, or a classification model constructed using a classifier.
  • VAD model voice emotion detection model
  • the audio data feature is no longer a single use of audio short-term energy and cepstrum features, but on this basis, the MFCC feature and the second-order difference feature of the audio information are further extracted to ensure that the audio information is in the frequency domain and time domain.
  • Completeness and coherence and then input the feature into a neural network model (for example, long and short-term memory network LSTM) for training, extract high-dimensional information and audio before and after related features, increase the recognition of noise and non-noise, and finally use a fully connected network As a classifier, it outputs the detection result of each frame.
  • a neural network model for example, long and short-term memory network LSTM
  • Step S3 includes:
  • the training steps of the neural network model and the fully connected network model include:
  • the optimizer selects Adam, the learning rate is e -3 , the number of training times is 100, and the batch size is 2048.
  • Perform training to extract high-dimensional information and audio signals of the audio signal The front and back correlation feature, the high-dimensional information is a highly abstract phoneme feature learned from a large number of samples by the neural network model, and the front and back correlation feature of the audio signal is obtained through the second-order difference feature of the audio signal.
  • the loss function of the neural network model is constructed by the following formula
  • y is the label of the sample
  • the positive class (human voice) is 1
  • the negative class (non-human voice) is 0
  • y′ is the output of the activation function, between 0-1
  • ⁇ is the output factor, ⁇ > 0.
  • step S6 after step S6, it further includes: smoothing the output result of the fully connected network model, and when detecting a frame that is a human voice or a non-human voice endpoint, then detecting the first setting before and after the frame A number of frame data, only when at least the second set number of detection results are consistent with the current frame detection results, the frame is finally determined to be human voice or non-human voice.
  • the detection result of at least 9 frames is consistent with the detection result of the current frame before it is finally determined that the frame is human voice or non-human voice.
  • the training steps of the neural network model and the fully connected network model further include:
  • the predicted value of the fully connected network model output is between 0-1, and 0 is non-human Voice, 1 is human voice, and the samples with the predicted value within the 0.5 setting range are difficult to classify samples.
  • the loss function is Modified based on the standard cross entropy loss, it can reduce the weight of easy-to-classify samples to make the model focus more on difficult-to-classify samples during training, improve the accuracy of the model for low-signal-to-noise sound recognition in non-stationary environments, and reduce the number of errors. Misjudgment of noise audio. Finally, perform a smooth operation on the output result of the model.
  • the model When the model detects that a certain frame is a human voice or non-human voice endpoint, it will detect the data of 10 frames before and after that frame. Only when at least 9 of the audio results of the frame are consistent with the detection result of the current frame At this time, the frame is finally determined to be human voice and non-human voice, in order to suppress some sudden changes in the information to cause the model to misjudge the result.
  • this application also provides a voice endpoint detection device for emotion recognition, including:
  • Acquisition module to collect audio signals
  • a signal processing module for processing the audio signals collected by the acquisition module, the processing operations include: adding pure noise segments and vocal noise segments in various scenarios to the audio signal and randomly setting the signal-to-noise ratio;
  • the first feature extraction module extracts the MFCC feature and the second-order difference feature of the audio signal processed by the signal processing module;
  • the second feature extraction module inputs the MFCC features extracted by the first feature extraction module and their second-order differential features into the neural network model, and extracts the high-dimensional information of the audio signal and the associated features of the audio signal before and after the high-dimensional information is a neural network
  • the highly abstract phoneme features learned by the model from the samples, and the associated features represent the temporal dynamic relationship of adjacent multiple frames of the audio signal;
  • the endpoint recognition module inputs the high-dimensional information and associated features of the audio signal extracted by the second feature extraction module into the fully connected network model to obtain the detection result of each frame of the audio signal, and the detection result includes human voice and non-human voice;
  • the segmentation module divides the audio signal into the human voice part and the non-human voice part according to the detection result of the audio signal.
  • it further includes an emotion recognition module, which inputs the audio signal of the human voice part into the voice emotion detection model, and outputs the result of voice endpoint detection for emotion recognition.
  • an emotion recognition module which inputs the audio signal of the human voice part into the voice emotion detection model, and outputs the result of voice endpoint detection for emotion recognition.
  • the endpoint recognition module performs a smooth operation on the output result of the fully connected network model, and when a frame that is a human voice or a non-human voice endpoint is detected, the first setting before and after the frame is detected For the number of frame data, only when at least the second set number of detection results are consistent with the detection result of the one frame, the one frame is determined to be human voice or non-human voice.
  • the first feature extraction module includes:
  • Pre-emphasis unit pre-emphasis the audio signal
  • Framing and windowing unit to frame and window the pre-emphasized audio signal
  • Transformation unit which performs fast Fourier transform on the framed and windowed audio signal
  • the smoothing processing unit smoothes the frequency spectrum of the audio signal after the fast Fourier transform through a triangular band-pass filter, and eliminates the effect of harmonics to highlight the formant of the original voice;
  • the MFCC obtaining unit calculates the logarithmic energy output by the triangular band-pass filter, and obtains the MFCC through the discrete cosine transform;
  • the second-order difference feature obtaining unit obtains the dynamic relationship between three adjacent frames through the relationship between the previous first-order difference and the latter first-order difference of the MFCC, so as to obtain the second-order difference feature.
  • it further includes a training module to train the neural network model and the fully connected network model.
  • the training module includes:
  • the noise adding unit obtains the human voice audio signal from the AISHELL data set, and adds a random combination of the pure noise segment and the human voice noise segment.
  • the AISHELL data set is an open source database of Beijing Hill Shell Technology Co., Ltd.;
  • the training set construction unit collects noise signals and human voice data in various scenarios and randomly sets the signal-to-noise ratio, thereby obtaining the training set of the human voice audio signals;
  • the feature extraction unit extracts the MFCC feature and its second-order difference feature of the audio signal in the training set
  • the training unit inputs the above-mentioned MFCC features and their second-order differential features into the neural network model, performs training, and extracts the high-dimensional information of the audio signal and the associated features before and after the audio signal.
  • the high-dimensional information is the neural network model learned from a large number of samples.
  • the highly abstract phoneme features of the audio signal are obtained through the second-order differential feature of the audio signal to obtain the before and after correlation characteristics of the audio signal; the high-dimensional information and the correlation characteristics of the audio signal are input into the fully connected network model to obtain the detection result of each frame of the audio signal;
  • the parameter update unit uses Focal loss as a loss function to iteratively update the parameters of the fully connected network model and the neural network model.
  • the loss function of the neural network model is constructed by
  • y is the label of the sample
  • y' is the output after the activation function, between 0-1
  • ⁇ is the output factor
  • ⁇ >0 is the label of the sample
  • Adam is selected as the optimizer of the neural network model, the learning rate is e -3 , the number of training times is 100, and the batch size is 2048.
  • the parameter update unit further modifies the weights of samples in the training set before each model parameter update, reduces the weights of easy-to-classify samples, and increases the weights of difficult-to-classify samples.
  • the predicted value output by the fully connected network model is 0. Between -1, 0 is non-human voice, 1 is human voice, and the sample with the predicted value within the set range of 0.5 is the difficult-to-classify sample.
  • the embodiment of the present application also proposes a computer-readable storage medium, which includes a voice endpoint detection program for emotion recognition, and the computer-readable storage medium may be non-volatile or It is volatile.
  • the voice endpoint detection program for emotion recognition is executed by the processor, the following steps are implemented:
  • processing operations on the audio signal including: adding pure noise segments and vocal noise segments in various scenarios to the audio signal and randomly setting a signal-to-noise ratio;
  • the audio signal is divided into a human voice part and a non-human voice part.
  • the specific implementation of the computer-readable storage medium of the present application is substantially the same as the specific implementation of the voice endpoint detection method and electronic device for emotion recognition, and will not be repeated here.
  • the neural network model and the fully connected network model can be synthesized into one, that is, the fully connected neural network model.
  • the fully connected neural network model may include one layer of LSTM and two layers of fully connected layer (FC).
  • the voice endpoint detection method, electronic device, and computer-readable storage medium for emotion recognition in the above embodiments are improved for the processing before voice emotion detection, mainly improving the low signal-to-noise ratio and non-stationary environment in traditional VAD Under the recognition of noise data, through data structure and model training, especially to enhance the training of difficult-to-classify data, it effectively suppresses the model’s misrecognition rate of noise or interference data, and the smooth operation of sudden signal also makes the sound cut the end point. More smooth and accurate. Because the human voice endpoint detection is added, the accuracy of voice emotion detection is effectively improved, the time and space consumption is reduced, and the working efficiency of the system is improved.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Hospice & Palliative Care (AREA)
  • General Health & Medical Sciences (AREA)
  • Psychiatry (AREA)
  • Child & Adolescent Psychology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Circuit For Audible Band Transducer (AREA)
  • Image Analysis (AREA)

Abstract

提供一种用于情绪识别的语音端点检测方法、装置、电子设备及存储介质，涉及人工智能，其中该方法包括：采集音频信号（S1）；对音频信号进行处理操作，包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比（S2）；提取处理后音频信号的MFCC特征及其二阶差分特征（S3）；将该特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征（S4）；将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，检测结果包括人声和非人声（S5）；根据音频信号的检测结果将音频信号分割成人声部分和非人声部分（S6）。该语音端点检测方法能够在低信噪比环境和非平稳环境下准确语音端点检测。

Description

用于情绪识别的语音端点检测方法、装置、电子设备及存储介质
本申请要求于2020年4月14日提交中国专利局、申请号为202010287911.2，发明名称为“用于情绪识别的语音端点检测方法、电子设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能的语音语义技术领域，更为具体地，涉及一种用于情绪识别的语音端点检测方法、装置、电子设备及存储介质。
背景技术
在自然语言处理过程中进行语音情绪识别前，需要准确识别出一段长音频中人声说话的端点位置，以便将环境噪声与说话人声进行分离，该技术即为语音端点检测(VAD)，是一种驱动性的语音信号处理技术。研究表明，若能准确识别说话人的起止位置并进行分割可以有效提升后续语音任务的准确率。发明人发现传统VAD技术主要基于音频的短时能量、过零率、倒谱特征或熵进行检测，这些方法原理简单，运算量小，但是当噪音和语音信号的时域和频域分布相似时，情绪识别效果比较差；在一些语音编码如G.729B和AMR-NB中也采用了VAD技术，这些方法同样在平稳噪声环境下准确率较高，但是在低信噪比环境和非平稳环境下检测效果就不理想。
发明内容
鉴于上述问题，本申请的目的是提供一种能够在低信噪比环境和非平稳环境下准确识别人声说话的端点位置的用于情绪识别的语音端点检测方法、装置、电子设备及存储介质。
为了实现上述目的，本申请提供一种用于情绪识别的语音端点检测方法，包括：
采集音频信号；
对音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
提取处理后的音频信号的MFCC特征及其二阶差分特征；
将上述MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从样本中学习到的高度抽象的音素特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
此外，为了实现上述目的，本申请还提供一种用于情绪识别的语音端点检测装置，包括：
采集模块，采集音频信号；
信号处理模块，对采集模块采集的音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
第一特征提取模块，提取信号处理模块处理后的音频信号的MFCC特征及其二阶差分特征；
第二特征提取模块，将上述第一特征提取模块提取的MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从样本中学习到的高度抽象的音素特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
端点识别模块，将第二特征提取模块提取的音频信号的高维信息及关联特征输入全连 接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
分割模块，根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
此外，为了实现上述目的，本申请还提供一种电子设备，所述电子设备包括存储器和处理器，所述存储器中存储有用于情绪识别的语音端点检测程序，所述用于情绪识别的语音端点检测程序被所述处理器执行时实现上述用于情绪识别的语音端点检测方法的步骤。此外，为了实现上述目的，本申请还提供一种计算机可读存储介质，所述计算机可读存储介质中包括有用于情绪识别的语音端点检测程序，所述用于情绪识别的语音端点检测程序被处理器执行时，实现上述用于情绪识别的语音端点检测方法的步骤。
本申请所述用于情绪识别的语音端点检测方法、装置、电子设备及存储介质对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比，将MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，将音频信号的高维信息及关联特征输入全连接网络模型，增强模型的在复杂多变环境下的鲁棒性和泛化能力，改善了传统VAD中对低信噪比和非平稳环境下噪声数据的辨识度。
附图说明
图1是本申请用于情绪识别的语音端点检测方法较佳实施例的应用环境示意图；
图2是图1中用于情绪识别的语音端点检测程序较佳实施例的模块示意图；
图3是本申请用于情绪识别的语音端点检测方法较佳实施例的流程图。
具体实施方式
应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
以下将结合附图对本申请的具体实施例进行详细描述。
本申请提供一种用于情绪识别的语音端点检测方法，应用于一种电子设备。参照图1所示，为本申请用于情绪识别的语音端点检测方法较佳实施例的应用环境示意图。
在本实施例中，电子设备1可以是服务器、手机、平板电脑、便携计算机、桌上型计算机等具有运算功能的终端客户端。
该电子设备1包括存储器11、处理器12、网络接口13及通信总线14。
存储器11包括至少一种类型的可读存储介质。所述至少一种类型的可读存储介质可为如闪存、硬盘、多媒体卡、卡型存储器等的非易失性存储介质。在一些实施例中，所述可读存储介质可以是所述电子设备1的内部存储单元，例如该电子设备1的硬盘。在另一些实施例中，所述可读存储介质也可以是所述电子设备1的外部存储器，例如所述电子设备1上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。
在本实施例中，所述存储器11的可读存储介质通常用于存储安装于所述电子设备1的用于情绪识别的语音端点检测程序10等。所述存储器11还可以用于暂时地存储已经输出或者将要输出的数据。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit，CPU)，微处理器或其他数据处理芯片，用于运行存储器11中存储的程序代码或处理数据，例如执行用于情绪识别的语音端点检测程序10等。
网络接口13可选地可以包括标准的有线接口、无线接口(如WI-FI接口)，通常用于在该电子设备1与其他电子客户端之间建立通信连接。
通信总线14用于实现这些组件之间的连接通信。
图1仅示出了具有组件11-14的电子设备1，但是应理解的是，并不要求实施所有示出的组件，可以替代的实施更多或者更少的组件。
可选地，该电子设备1还可以包括用户接口，用户接口可以包括输入单元比如键盘(Keyboard)、语音输入装置比如麦克风(microphone)等具有语音识别功能的客户端、语音输 出装置比如音响、耳机等，可选地用户接口还可以包括标准的有线接口、无线接口。
可选地，该电子设备1还可以包括显示器，显示器也可以称为显示屏或显示单元。
在一些实施例中可以是LED显示器、液晶显示器、触控式液晶显示器以及有机发光二极管(Organic Light-Emitting Diode，OLED)触摸器等。显示器用于显示在电子设备1中处理的信息以及用于显示可视化的用户界面。
可选地，该电子设备1还包括触摸传感器。所述触摸传感器所提供的供用户进行触摸操作的区域称为触控区域。此外，这里所述的触摸传感器可以为电阻式触摸传感器、电容式触摸传感器等。而且，所述触摸传感器不仅包括接触式的触摸传感器，也可包括接近式的触摸传感器等。此外，所述触摸传感器可以为单个传感器，也可以为例如阵列布置的多个传感器。
可选地，该电子设备1还可以包括逻辑门电路，传感器、音频电路等等，在此不再赘述。
在图1所示的装置实施例中，作为一种计算机存储介质的存储器11中可以包括操作系统以及用于情绪识别的语音端点检测程序10；处理器12执行存储器11中存储的用于情绪识别的语音端点检测程序10时实现如下步骤：
采集音频信号；
对音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
提取处理后的音频信号的MFCC特征及其二阶差分特征，所述MFCC特征为梅尔倒谱系数(Mel-scale Frequency Cepstral Coefficients，简称MFCC)是在Mel标度频率域提取出来的倒谱参数，Mel标度描述了人耳频率的非线性特性，所述MFCC特征的二阶差分特征是音频信号相邻三帧之间的动态关系；
将上述MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从大量样本中学习到的高度抽象的音素特征，所述音素特征是指根据音素时长信息表达的词性特征、上下文信息和音调特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
根据音频信号的检测结果，将提取的音频信号分割成人声部分和非人声部分。
在其他实施例中，所述用于情绪识别的语音端点检测程序10还可以被分割为一个或者多个模块，一个或者多个模块被存储于存储器11中，并由处理器12执行，以完成本申请。本申请所称的模块是指能够完成特定功能的一系列计算机程序指令段。参照图2所示，为图1中用于情绪识别的语音端点检测程序10较佳实施例的功能模块图。所述用于情绪识别的语音端点检测程序10可以被分割为采集模块110、信号处理模块120、第一特征提取模块130、第二特征提取模块140、端点识别模块150和分割模块160，其中：
采集模块110，采集音频信号；
信号处理模块120，对采集模块110采集的音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
第一特征提取模块130，提取信号处理模块120处理后的音频信号的MFCC特征及其二阶差分特征；
第二特征提取模块140，将上述第一特征提取模块130提取的MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征；
端点识别模块150，将第二特征提取模块140提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
分割模块160，根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
优选地，还包括情绪识别模块170，将人声部分的音频信号输入语音情绪检测模型，输出用于情绪识别的语音端点检测的结果。
在一个可选实施例中，上述端点识别模块150对全连接网络模型的输出结果进行平滑操作，当检测到为人声或非人声端点的一帧时，再检测所述一帧的前后第一设定数量的帧数据，只有当其中至少第二设定数量的检测结果与所述一帧的检测结果一致时，才最终确定所述一帧为人声或非人声。
在一个可选实施例中，所述第一特征提取模块130包括：
预加重单元，对音频信号进行预加重；
分帧加窗单元，对预加重后的音频信号进行分帧和加窗；
变换单元，对分帧和加窗后的音频信号进行快速傅里叶变换；
平滑处理单元，经过三角带通滤波器对经过快速傅立叶变换后的音频信号的频谱进行平滑化，并消除谐波的作用，突显原先语音的共振峰；
MFCC获得单元，计算三角带通滤波器输出的对数能量，经离散余弦变换得到MFCC；
二阶差分特征获得单元，通过MFCC前一阶差分与后一阶差分之间的关系，获得相邻三帧之间的动态关系，从而获得二阶差分特征。
在一个实施例中，还包括训练模块，对所述神经网络模型和全连接网络模型进行训练，包括：
噪声添加单元，从AISHELL数据集获得人声音频信号，加入纯噪声段和人声噪声段随机组合，所述AISHELL数据集为北京希尔贝壳科技有限公司的开源数据库；
训练集构建单元，进行多种场景下的噪声信号与人声数据的采集并随机设置信噪比，从而获得所述人声音频信号的训练集；
特征提取单元，提取训练集中音频信号的MFCC特征及其二阶差分特征；
训练单元，将上述MFCC特征及其二阶差分特征输入神经网络模型，进行训练，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从大量样本中学习到的高度抽象的音素特征，通过音频信号的二阶差分特征获得所述音频信号前后关联特征；将音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果；
参数更新单元，使用Focal loss作为损失函数对全连接网络模型和神经网络模型的参数进行迭代更新。
优选地，上述参数更新单元还在每次模型参数更新之前，修改训练集中样本的权重，减少易分类样本的权重，增加难分类样本的权重。
此外，本申请还提供一种用于情绪识别的语音端点检测方法。参照图3所示，为本申请用于情绪识别的语音端点检测较佳实施例的流程图。该方法可以由一个装置执行，该装置可以由软件和/或硬件实现。
在本实施例中，用于情绪识别的语音端点检测方法包括：
步骤S1，采集音频信号；
步骤S2，对音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比，所述多种场景包括办公室噪声集、手录实验室噪声集、电话催收噪声集和面审数据等，例如，在20dB-40dB范围内设置信噪比，根据设置的信噪比获得音频信号和噪声信号的比例，噪声信号由纯噪声段和人声噪声段构成，噪声信号的功率确定，调整纯噪声段和人声噪声段的混合比例，并与音频信号相加堆叠，获得最接近人声(不淹没音频信号)的混合比例；
步骤S3，提取处理后的音频信号的MFCC特征及其二阶差分特征，所述MFCC特征 及其二阶差分特征的提取方法包括：将添加噪声后的音频信号经过高通滤波器进行预加重，对预加重后的音频信号进行分帧和加窗，对分帧加窗后的音频信号进行快速傅立叶变换，经过快速傅立叶变换后的音频信号通过Mel频率滤波器组进行平滑处理，计算每个滤波器组输出的对数能量，经离散余弦变换(DCT)得到MFCC(音频信号的MFCC特征)，提取MFCC的差分谱，获得其二阶差分特征；
步骤S4，将上述MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征；
步骤S5，将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声，例如采用人声为1，非人声为0，将音频信号变成二值序列；
步骤S6，根据音频信号的检测结果将音频信号分割成人声部分和非人声部分，例如，将音频信号中检测结果为1的帧进行相加堆叠，将检测结果为0的帧进行相加堆叠，实现人声部分和非人声部分的分割。
优选地，步骤S6之后还包括：将人声部分的音频信号输入语音情绪检测模型，输出用于情绪识别的语音端点检测的结果，例如，将语音情绪分成快乐、兴趣、厌恶、恐惧、痛苦(悲伤)和愤怒等，所述语音情绪检测模型可以是基于神经网络构建的分类模型，也可以是采用分类器构建的分类模型。
在实时语音情绪检测系统中，长时间的静音或者噪声一方面增加了网络通信的代价，另一方面也会使情绪检测系统产生错误判断。准确识别音频中人声起止位置，不仅可以提高模型运行效率，也可以提升模型的稳定性和准确性。
作为进行语音情绪检测前的音频预处理操作，保证数据特征的一致性可以有效提升情绪检测的准确率，所以在设计语音情绪检测模型(VAD模型)时，采用与情绪检测预训练模型相同的数据源来构造训练集。另外，为了增强模型的在复杂多变环境下的鲁棒性和泛化能力，进行多种场景下的噪声数据与人声数据的采集并随机设置信噪比保证数据的广泛性。
音频数据特征不再是单一的使用音频短时能量和倒谱特征等，而是在此基础上进一步提取音频信息的MFCC特征及其二阶差分特征，保证音频在频域和时域上信息的完整性和连贯性，之后将该特征输入神经网络模型(例如，长短期记忆网络LSTM)进行训练，提取高维信息及音频前后关联特征，增加噪声和非噪声的识别度，最后使用全连接网络作为分类器输出每一帧的检测结果。
步骤S3包括：
对音频信号进行预加重；
对预加重后的音频信号进行分帧和加窗；
对分帧和加窗后的音频信号进行快速傅里叶变换；
经过三角带通滤波器对经过快速傅立叶变换后的音频信号的频谱进行平滑化，并消除谐波的作用，突显原先语音的共振峰；
计算三角带通滤波器输出的对数能量，经离散余弦变换(DCT)得到MFCC；
通过MFCC前一阶差分与后一阶差分之间的关系，获得相邻三帧之间的动态关系，从而获得二阶差分特征。
在一个实施例中，所述神经网络模型和全连接网络模型的训练步骤包括：
从AISHELL数据集获得人声音频信号，加入纯噪声段和人声噪声段随机组合；
进行多种场景下的噪声信号与人声数据的采集并随机设置信噪比，从而获得所述人声音频信号的训练集；
提取训练集中音频信号的MFCC特征及其二阶差分特征；
将上述MFCC特征及其二阶差分特征输入神经网络模型，优化器选用Adam，学习率为e -3，训练次数为100，批尺寸为2048，进行训练，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从大量样本中学习到的高度抽象的音素特征，通过音频信号的二阶差分特征获得所述音频信号前后关联特征。
将音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果；
使用Focal loss作为损失函数对全连接网络模型和神经网络模型的参数进行迭代更新，优选地，通过下式构建神经网络模型的损失函数
Figure PCTCN2020104783-appb-000001
其中，y是样本的标签，正类(人声)为1，负类(非人声)为0，y′是经过激活函数的输出，在0-1之间；γ为输出因子，γ>0。
普通的交叉熵对于正样本而言，输出概率越大损失越小。对于负样本而言，输出概率越小则损失越小。此时的损失函数在大量简单样本的迭代过程中比较缓慢且可能无法优化至最优，在原有的基础上加了一个因子，其中，γ>0使得减少易分类样本的损失。使得更关注于困难的、错分的样本。
在一个实施例中，在步骤S6之后还包括：对全连接网络模型的输出结果进行平滑操作，当检测到为人声或非人声端点的一帧时，再检测该帧的前后第一设定数量的帧数据，只有当其中至少第二设定数量的检测结果与当前帧检测结果一致时，才最终确定该帧为人声或非人声，例如，检测该帧的前后10帧数据，只有当其中至少9帧检测结果与当前帧检测结果一致才最终确定该帧为人声或非人声。
优选地，神经网络模型和全连接网络模型的训练步骤还包括：
在每次模型参数更新之前，修改训练集中样本的权重，减少易分类样本的权重，增加难分类样本的权重，例如，全连接网络模型输出的预测值在0-1之间，0为非人声，1为人声，预测值在0.5设定范围内的样本为难分类样本。
在平稳环境下，人声和噪声具有较高的辨识度，但是当环境为非平稳情况时，人声和非人声就难以辨识，所以在这里使用Focal loss作为损失函数，该损失函数是在标准交叉熵损失基础上修改得到，可以通过减少易分类样本的权重，使模型在训练时更加专注于难分类样本，提高模型在非平稳环境下对低信噪比声音识别的准确度，减少对噪声音频的误判。最后对模型的输出结果进行平滑操作，当模型检测到某一帧为人声或非人声端点时，再检测该帧的前后10帧数据，只有当其中至少9帧音频结果与当前帧检测结果一致时，才最终确定该帧为人声和非人声，以抑制某些突变信息造成模型对结果的误判。
此外，本申请还提供一种用于情绪识别的语音端点检测装置，包括：
采集模块，采集音频信号；
信号处理模块，对采集模块采集的音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
第一特征提取模块，提取信号处理模块处理后的音频信号的MFCC特征及其二阶差分特征；
第二特征提取模块，将上述第一特征提取模块提取的MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从样本中学习到的高度抽象的音素特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
端点识别模块，将第二特征提取模块提取的音频信号的高维信息及关联特征输入全连 接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
分割模块，根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
优选地，还包括情绪识别模块，所述情绪识别模块将人声部分的音频信号输入语音情绪检测模型，输出用于情绪识别的语音端点检测的结果。
在一个实施例中，所述端点识别模块对全连接网络模型的输出结果进行平滑操作，当检测到为人声或非人声端点的一帧时，再检测所述一帧的前后第一设定数量的帧数据，只有当其中至少第二设定数量的检测结果与所述一帧的检测结果一致时，才确定所述一帧为人声或非人声。
在一个实施例中，所述第一特征提取模块包括：
预加重单元，对音频信号进行预加重；
分帧加窗单元，对预加重后的音频信号进行分帧和加窗；
变换单元，对分帧和加窗后的音频信号进行快速傅里叶变换；
平滑处理单元，经过三角带通滤波器对经过快速傅立叶变换后的音频信号的频谱进行平滑化，并消除谐波的作用，突显原先语音的共振峰；
MFCC获得单元，计算三角带通滤波器输出的对数能量，经离散余弦变换得到MFCC；
二阶差分特征获得单元，通过MFCC前一阶差分与后一阶差分之间的关系，获得相邻三帧之间的动态关系，从而获得二阶差分特征。
在一个实施例中，还包括训练模块，对所述神经网络模型和全连接网络模型进行训练。
优选地，所述训练模块包括：
噪声添加单元，从AISHELL数据集获得人声音频信号，加入纯噪声段和人声噪声段随机组合，所述AISHELL数据集为北京希尔贝壳科技有限公司的开源数据库；
训练集构建单元，进行多种场景下的噪声信号与人声数据的采集并随机设置信噪比，从而获得所述人声音频信号的训练集；
特征提取单元，提取训练集中音频信号的MFCC特征及其二阶差分特征；
训练单元，将上述MFCC特征及其二阶差分特征输入神经网络模型，进行训练，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从大量样本中学习到的高度抽象的音素特征，通过音频信号的二阶差分特征获得所述音频信号前后关联特征；将音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果；
参数更新单元，使用Focal loss作为损失函数对全连接网络模型和神经网络模型的参数进行迭代更新。
优选地，所述神经网络模型的损失函数通过下式构建
Figure PCTCN2020104783-appb-000002
其中，y是样本的标签，y′是经过激活函数的输出，在0-1之间，γ为输出因子，γ>0。
优选地，所述神经网络模型的优化器选用Adam，学习率为e -3，训练次数为100，批尺寸为2048。
优选地，所述参数更新单元还在每次模型参数更新之前，修改训练集中样本的权重，减少易分类样本的权重，增加难分类样本的权重，所述全连接网络模型输出的预测值在0-1之间，0为非人声，1为人声，预测值在0.5的设定范围内的样本为难分类样本。
此外，本申请实施例还提出一种计算机可读存储介质，所述计算机可读存储介质中包括用于情绪识别的语音端点检测程序，所述计算机可读存储介质可以是非易失性，也可以是易失性，所述用于情绪识别的语音端点检测程序被处理器执行时实现如下步骤：
采集音频信号；
对音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
提取处理后的音频信号的MFCC特征及其二阶差分特征；
将上述MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征；
将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
本申请之计算机可读存储介质的具体实施方式与上述用于情绪识别的语音端点检测方法、电子设备的具体实施方式大致相同，在此不再赘述。
在上述各实施例中，神经网络模型和全连接网络模型可以合成为一个，即全连接神经网络模型，所述全连接神经网络模型可以包括一层LSTM和两层全连接层(FC)。
上述各实施例中的用于情绪识别的语音端点检测方法、电子设备和计算机可读存储介质针对语音情绪检测前的处理进行了改进，主要改善了传统VAD中对低信噪比和非平稳环境下噪声数据的辨识度，通过数据构造和模型训练的方式，尤其增强对难分类数据的训练，有效抑制了模型对噪声或干扰数据的误识别率，对于突变信号的平滑操作也使得声音切割端点更加平滑准确。因为加入人声端点检测，所以有效提高了语音情绪检测的准确性，减少了时间和空间的消耗，提升了系统的工作效率。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、装置、物品或者方法不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、装置、物品或者方法所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、装置、物品或者方法中还存在另外的相同要素。
上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端客户端(可以是手机，计算机，服务器，或者网络客户端等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种用于情绪识别的语音端点检测方法，其中，包括：
    采集音频信号；
    对音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
    提取处理后的音频信号的MFCC特征及其二阶差分特征；
    将上述MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从样本中学习到的高度抽象的音素特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
    将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
    根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
  2. 根据权利要求1所述的用于情绪识别的语音端点检测方法，其中，所述提取处理后的音频信号的MFCC特征及其二阶差分特征的步骤包括：
    对音频信号进行预加重；
    对预加重后的音频信号进行分帧和加窗；
    对分帧和加窗后的音频信号进行快速傅里叶变换；
    经过三角带通滤波器对经过快速傅立叶变换后的音频信号的频谱进行平滑化，并消除谐波的作用，突显原先语音的共振峰；
    计算三角带通滤波器输出的对数能量，经离散余弦变换得到MFCC；
    通过MFCC前一阶差分与后一阶差分之间的关系，获得相邻三帧之间的动态关系，从而获得二阶差分特征。
  3. 根据权利要求1所述的用于情绪识别的语音端点检测方法，其中，所述神经网络模型和全连接网络模型的训练步骤包括：
    从AISHELL数据集获得人声音频信号，加入纯噪声段和人声噪声段随机组合；
    进行多种场景下的噪声信号与人声数据的采集并随机设置信噪比，从而获得所述人声音频信号的训练集；
    提取训练集中音频信号的MFCC特征及其二阶差分特征；
    将上述MFCC特征及其二阶差分特征输入神经网络模型，进行训练，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从大量样本中学习到的高度抽象的音素特征，通过音频信号的二阶差分特征获得所述音频信号前后关联特征；
    将音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果；
    使用Focal loss作为损失函数对全连接网络模型和神经网络模型的参数进行迭代更新。
  4. 根据权利要求3所述的用于情绪识别的语音端点检测方法，其中，所述神经网络模型的损失函数通过下式构建
    Figure PCTCN2020104783-appb-100001
    其中，y是样本的标签，y′是经过激活函数的输出，在0-1之间，γ为输出因子，γ>0。
  5. 根据权利要求3所述的用于情绪识别的语音端点检测方法，其中，所述神经网络模型的优化器选用Adam，学习率为e -3，训练次数为100，批尺寸为2048。
  6. 根据权利要求3所述的用于情绪识别的语音端点检测方法，其中，所述神经网络模 型和全连接网络模型的训练步骤还包括：
    在每次模型参数更新之前，修改训练集中样本的权重，减少易分类样本的权重，增加难分类样本的权重，所述全连接网络模型输出的预测值在0-1之间，0为非人声，1为人声，预测值在0.5的设定范围内的样本为难分类样本。
  7. 根据权利要求1所述的用于情绪识别的语音端点检测方法，其中，还包括：
    将人声部分的音频信号输入语音情绪检测模型，输出语音情绪识别的结果。
  8. 根据权利要求1所述的用于情绪识别的语音端点检测方法，其中，还包括：对全连接网络模型的输出结果进行平滑操作，当检测到为人声或非人声端点的一帧时，再检测所述一帧的前后第一设定数量的帧数据，只有当其中至少第二设定数量的检测结果与所述一帧的检测结果一致时，才最终确定所述一帧为人声或非人声。
  9. 一种用于情绪识别的语音端点检测装置，其中，包括：
    采集模块，采集音频信号；
    信号处理模块，对采集模块采集的音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
    第一特征提取模块，提取信号处理模块处理后的音频信号的MFCC特征及其二阶差分特征；
    第二特征提取模块，将上述第一特征提取模块提取的MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从样本中学习到的高度抽象的音素特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
    端点识别模块，将第二特征提取模块提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
    分割模块，根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
  10. 一种电子设备，其中，包括存储器和处理器，所述存储器中存储有用于情绪识别的语音端点检测程序，所述用于情绪识别的语音端点检测程序被所述处理器执行时实现如下步骤：
    采集音频信号；
    对音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
    提取处理后的音频信号的MFCC特征及其二阶差分特征；
    将上述MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从样本中学习到的高度抽象的音素特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
    将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
    根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
  11. 根据权利要求10所述的电子设备，其中，所述提取处理后的音频信号的MFCC特征及其二阶差分特征的步骤包括：
    对音频信号进行预加重；
    对预加重后的音频信号进行分帧和加窗；
    对分帧和加窗后的音频信号进行快速傅里叶变换；
    经过三角带通滤波器对经过快速傅立叶变换后的音频信号的频谱进行平滑化，并消除谐波的作用，突显原先语音的共振峰；
    计算三角带通滤波器输出的对数能量，经离散余弦变换得到MFCC；
    通过MFCC前一阶差分与后一阶差分之间的关系，获得相邻三帧之间的动态关系，从而获得二阶差分特征。
  12. 根据权利要求10所述的电子设备，其中，所述神经网络模型和全连接网络模型的训练步骤包括：
    从AISHELL数据集获得人声音频信号，加入纯噪声段和人声噪声段随机组合；
    进行多种场景下的噪声信号与人声数据的采集并随机设置信噪比，从而获得所述人声音频信号的训练集；
    提取训练集中音频信号的MFCC特征及其二阶差分特征；
    将上述MFCC特征及其二阶差分特征输入神经网络模型，进行训练，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从大量样本中学习到的高度抽象的音素特征，通过音频信号的二阶差分特征获得所述音频信号前后关联特征；
    将音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果；
    使用Focal loss作为损失函数对全连接网络模型和神经网络模型的参数进行迭代更新。
  13. 根据权利要求12所述的电子设备，其中，所述神经网络模型的损失函数通过下式构建
    Figure PCTCN2020104783-appb-100002
    其中，y是样本的标签，y′是经过激活函数的输出，在0-1之间，γ为输出因子，γ>0。
  14. 根据权利要求10所述的电子设备，其中，所述用于情绪识别的语音端点检测程序被所述处理器执行时还实现如下步骤：
    将人声部分的音频信号输入语音情绪检测模型，输出语音情绪识别的结果。
  15. 根据权利要求10所述的电子设备，其中，所述用于情绪识别的语音端点检测程序被所述处理器执行时还实现如下步骤：对全连接网络模型的输出结果进行平滑操作，当检测到为人声或非人声端点的一帧时，再检测所述一帧的前后第一设定数量的帧数据，只有当其中至少第二设定数量的检测结果与所述一帧的检测结果一致时，才最终确定所述一帧为人声或非人声。
  16. 一种计算机可读存储介质，其中，所述计算机可读存储介质中包括有用于情绪识别的语音端点检测程序，所述用于情绪识别的语音端点检测程序被处理器执行时，实现如下步骤：
    采集音频信号；
    对音频信号进行处理操作，所述处理操作包括：对音频信号加入多种场景下的纯噪声段和人声噪声段并随机设置信噪比；
    提取处理后的音频信号的MFCC特征及其二阶差分特征；
    将上述MFCC特征及其二阶差分特征输入神经网络模型，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从样本中学习到的高度抽象的音素特征，所述关联特征表示音频信号相邻多帧的时序动态关系；
    将提取的音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果，所述检测结果包括人声和非人声；
    根据音频信号的检测结果，将音频信号分割成人声部分和非人声部分。
  17. 根据权利要求16所述的计算机可读存储介质，其中，所述提取处理后的音频信号的MFCC特征及其二阶差分特征的步骤包括：
    对音频信号进行预加重；
    对预加重后的音频信号进行分帧和加窗；
    对分帧和加窗后的音频信号进行快速傅里叶变换；
    经过三角带通滤波器对经过快速傅立叶变换后的音频信号的频谱进行平滑化，并消除谐波的作用，突显原先语音的共振峰；
    计算三角带通滤波器输出的对数能量，经离散余弦变换得到MFCC；
    通过MFCC前一阶差分与后一阶差分之间的关系，获得相邻三帧之间的动态关系，从而获得二阶差分特征。
  18. 根据权利要求16所述的计算机可读存储介质，其中，所述神经网络模型和全连接网络模型的训练步骤包括：
    从AISHELL数据集获得人声音频信号，加入纯噪声段和人声噪声段随机组合；
    进行多种场景下的噪声信号与人声数据的采集并随机设置信噪比，从而获得所述人声音频信号的训练集；
    提取训练集中音频信号的MFCC特征及其二阶差分特征；
    将上述MFCC特征及其二阶差分特征输入神经网络模型，进行训练，提取音频信号的高维信息及音频信号前后关联特征，所述高维信息是神经网络模型从大量样本中学习到的高度抽象的音素特征，通过音频信号的二阶差分特征获得所述音频信号前后关联特征；
    将音频信号的高维信息及关联特征输入全连接网络模型，获得音频信号每一帧的检测结果；
    使用Focal loss作为损失函数对全连接网络模型和神经网络模型的参数进行迭代更新。
  19. 根据权利要求16所述的计算机可读存储介质，其中，所述用于情绪识别的语音端点检测程序被处理器执行时，还实现如下步骤：
    将人声部分的音频信号输入语音情绪检测模型，输出语音情绪识别的结果。
  20. 根据权利要求16所述的计算机可读存储介质，其中，所述用于情绪识别的语音端点检测程序被处理器执行时，还实现如下步骤：对全连接网络模型的输出结果进行平滑操作，当检测到为人声或非人声端点的一帧时，再检测所述一帧的前后第一设定数量的帧数据，只有当其中至少第二设定数量的检测结果与所述一帧的检测结果一致时，才最终确定所述一帧为人声或非人声。
PCT/CN2020/104783 2020-04-14 2020-07-27 用于情绪识别的语音端点检测方法、装置、电子设备及存储介质 WO2021208287A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010287911.2A CN111179975B (zh) 2020-04-14 2020-04-14 用于情绪识别的语音端点检测方法、电子设备及存储介质
CN202010287911.2 2020-04-14

Publications (1)

Publication Number Publication Date
WO2021208287A1 true WO2021208287A1 (zh) 2021-10-21

Family

ID=70656748

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/104783 WO2021208287A1 (zh) 2020-04-14 2020-07-27 用于情绪识别的语音端点检测方法、装置、电子设备及存储介质

Country Status (2)

Country Link
CN (1) CN111179975B (zh)
WO (1) WO2021208287A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114299950A (zh) * 2021-12-30 2022-04-08 北京字跳网络技术有限公司 一种字幕生成方法、装置及设备
CN114399005A (zh) * 2022-03-10 2022-04-26 深圳市声扬科技有限公司 一种活体检测模型的训练方法、装置、设备及存储介质
CN114598767A (zh) * 2022-01-11 2022-06-07 浙江大学 基于惯性测量单元的跨设备手机语音信息恢复方法及系统
CN115171700A (zh) * 2022-06-13 2022-10-11 武汉大学 一种基于脉冲神经网络的声纹识别语音助手方法
CN115883753A (zh) * 2022-11-04 2023-03-31 网易（杭州）网络有限公司 视频的生成方法、装置、计算设备及存储介质
CN117475360A (zh) * 2023-12-27 2024-01-30 南京纳实医学科技有限公司 基于改进型mlstm-fcn的音视频特点的生物体征提取与分析方法

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111179975B (zh) * 2020-04-14 2020-08-04 深圳壹账通智能科技有限公司 用于情绪识别的语音端点检测方法、电子设备及存储介质
CN111986686B (zh) * 2020-07-09 2023-01-03 厦门快商通科技股份有限公司 短时语音信噪比估算方法、装置、设备及存储介质
CN112102846B (zh) * 2020-09-04 2021-08-17 腾讯科技（深圳）有限公司 音频处理方法、装置、电子设备以及存储介质
CN112215927B (zh) * 2020-09-18 2023-06-23 腾讯科技（深圳）有限公司 人脸视频的合成方法、装置、设备及介质
CN112102851B (zh) * 2020-11-17 2021-04-13 深圳壹账通智能科技有限公司 语音端点检测方法、装置、设备及计算机可读存储介质
CN112509601B (zh) * 2020-11-18 2022-09-06 中电海康集团有限公司 一种音符起始点检测方法及系统
CN112951259A (zh) * 2021-03-01 2021-06-11 杭州网易云音乐科技有限公司 音频降噪方法、装置、电子设备及计算机可读存储介质
CN113131965B (zh) * 2021-04-16 2023-11-07 成都天奥信息科技有限公司 一种民航甚高频地空通信电台遥控装置及人声判别方法
CN113113001A (zh) * 2021-04-20 2021-07-13 深圳市友杰智新科技有限公司 人声激活检测方法、装置、计算机设备和存储介质
CN112992191B (zh) * 2021-05-12 2021-11-05 北京世纪好未来教育科技有限公司 语音端点检测方法、装置、电子设备及可读存储介质
CN113327630B (zh) * 2021-05-27 2023-05-09 平安科技（深圳）有限公司 语音情绪识别方法、装置、设备及存储介质
CN113572908A (zh) * 2021-06-16 2021-10-29 云茂互联智能科技(厦门)有限公司 一种VoIP通话中降噪的方法、装置及系统
CN113724720B (zh) * 2021-07-19 2023-07-11 电信科学技术第五研究所有限公司 一种基于神经网络和mfcc的嘈杂环境下非人声语音过滤方法
CN113555028A (zh) * 2021-07-19 2021-10-26 首约科技(北京)有限公司 一种用于车联网语音降噪的处理方法
CN114333912B (zh) * 2021-12-15 2023-08-29 北京百度网讯科技有限公司 语音激活检测方法、装置、电子设备和存储介质
CN115065912B (zh) * 2022-06-22 2023-04-25 广东帝比电子科技有限公司 基于声纹筛技术的对音箱能量进行筛选的反馈抑制装置
CN115264408A (zh) * 2022-06-23 2022-11-01 大连理工大学 一种基于音频处理的管道泄漏检测方法及系统
CN115273915A (zh) * 2022-07-29 2022-11-01 歌尔科技有限公司 语音激活检测方法、装置、终端设备以及计算机存储介质

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103021406A (zh) * 2012-12-18 2013-04-03 台州学院 基于压缩感知的鲁棒性语音情感识别方法
CN106340309A (zh) * 2016-08-23 2017-01-18 南京大空翼信息技术有限公司 一种基于深度学习的狗叫情感识别方法及装置
CN109767781A (zh) * 2019-03-06 2019-05-17 哈尔滨工业大学（深圳） 基于超高斯先验语音模型与深度学习的语音分离方法、系统及存储介质
CN110491416A (zh) * 2019-07-26 2019-11-22 广东工业大学 一种基于lstm和sae的电话语音情感分析与识别方法
CN110706694A (zh) * 2019-09-26 2020-01-17 成都数之联科技有限公司 一种基于深度学习的语音端点检测方法及系统
US20200074997A1 (en) * 2018-08-31 2020-03-05 CloudMinds Technology, Inc. Method and system for detecting voice activity in noisy conditions
CN110991652A (zh) * 2019-12-02 2020-04-10 北京迈格威科技有限公司 神经网络模型训练方法、装置及电子设备
CN111179975A (zh) * 2020-04-14 2020-05-19 深圳壹账通智能科技有限公司 用于情绪识别的语音端点检测方法、电子设备及存储介质

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105118502B (zh) * 2015-07-14 2017-05-10 百度在线网络技术（北京）有限公司 语音识别系统的端点检测方法及系统
CN105261357B (zh) * 2015-09-15 2016-11-23 百度在线网络技术（北京）有限公司 基于统计模型的语音端点检测方法及装置
CN105321525B (zh) * 2015-09-30 2019-02-22 北京邮电大学 一种降低voip通信资源开销的系统和方法
GB2561408A (en) * 2017-04-10 2018-10-17 Cirrus Logic Int Semiconductor Ltd Flexible voice capture front-end for headsets
US11276390B2 (en) * 2018-03-22 2022-03-15 Casio Computer Co., Ltd. Audio interval detection apparatus, method, and recording medium to eliminate a specified interval that does not represent speech based on a divided phoneme
CN110910905B (zh) * 2018-09-18 2023-05-02 京东科技控股股份有限公司 静音点检测方法及装置、存储介质、电子设备
CN109887511A (zh) * 2019-04-24 2019-06-14 武汉水象电子科技有限公司 一种基于级联dnn的语音唤醒优化方法

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103021406A (zh) * 2012-12-18 2013-04-03 台州学院 基于压缩感知的鲁棒性语音情感识别方法
CN106340309A (zh) * 2016-08-23 2017-01-18 南京大空翼信息技术有限公司 一种基于深度学习的狗叫情感识别方法及装置
US20200074997A1 (en) * 2018-08-31 2020-03-05 CloudMinds Technology, Inc. Method and system for detecting voice activity in noisy conditions
CN109767781A (zh) * 2019-03-06 2019-05-17 哈尔滨工业大学（深圳） 基于超高斯先验语音模型与深度学习的语音分离方法、系统及存储介质
CN110491416A (zh) * 2019-07-26 2019-11-22 广东工业大学 一种基于lstm和sae的电话语音情感分析与识别方法
CN110706694A (zh) * 2019-09-26 2020-01-17 成都数之联科技有限公司 一种基于深度学习的语音端点检测方法及系统
CN110991652A (zh) * 2019-12-02 2020-04-10 北京迈格威科技有限公司 神经网络模型训练方法、装置及电子设备
CN111179975A (zh) * 2020-04-14 2020-05-19 深圳壹账通智能科技有限公司 用于情绪识别的语音端点检测方法、电子设备及存储介质

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114299950A (zh) * 2021-12-30 2022-04-08 北京字跳网络技术有限公司 一种字幕生成方法、装置及设备
CN114299950B (zh) * 2021-12-30 2023-07-14 北京字跳网络技术有限公司 一种字幕生成方法、装置及设备
CN114598767A (zh) * 2022-01-11 2022-06-07 浙江大学 基于惯性测量单元的跨设备手机语音信息恢复方法及系统
CN114399005A (zh) * 2022-03-10 2022-04-26 深圳市声扬科技有限公司 一种活体检测模型的训练方法、装置、设备及存储介质
CN115171700A (zh) * 2022-06-13 2022-10-11 武汉大学 一种基于脉冲神经网络的声纹识别语音助手方法
CN115171700B (zh) * 2022-06-13 2024-04-26 武汉大学 一种基于脉冲神经网络的声纹识别语音助手方法
CN115883753A (zh) * 2022-11-04 2023-03-31 网易（杭州）网络有限公司 视频的生成方法、装置、计算设备及存储介质
CN117475360A (zh) * 2023-12-27 2024-01-30 南京纳实医学科技有限公司 基于改进型mlstm-fcn的音视频特点的生物体征提取与分析方法
CN117475360B (zh) * 2023-12-27 2024-03-26 南京纳实医学科技有限公司 基于改进型mlstm-fcn的音视频特点的生物特征提取与分析方法

Also Published As

Publication number Publication date
CN111179975B (zh) 2020-08-04
CN111179975A (zh) 2020-05-19

Similar Documents

Publication Publication Date Title
WO2021208287A1 (zh) 用于情绪识别的语音端点检测方法、装置、电子设备及存储介质
CN109243491B (zh) 在频谱上对语音进行情绪识别的方法、系统及存储介质
CN112259106B (zh) 声纹识别方法、装置、存储介质及计算机设备
WO2018227781A1 (zh) 语音识别方法、装置、计算机设备及存储介质
WO2018227780A1 (zh) 语音识别方法、装置、计算机设备及存储介质
Li et al. Spatiotemporal and frequential cascaded attention networks for speech emotion recognition
CN110457432A (zh) 面试评分方法、装置、设备及存储介质
WO2020034628A1 (zh) 口音识别方法、装置、计算机装置及存储介质
Zvarevashe et al. Recognition of speech emotion using custom 2D-convolution neural network deep learning algorithm
CN113327626A (zh) 语音降噪方法、装置、设备及存储介质
CN110136726A (zh) 一种语音性别的估计方法、装置、系统及存储介质
WO2023279691A1 (zh) 语音分类方法、模型训练方法及装置、设备、介质和程序
Mahesha et al. LP-Hillbert transform based MFCC for effective discrimination of stuttering dysfluencies
Pao et al. A study on the search of the most discriminative speech features in the speaker dependent speech emotion recognition
Shah et al. Speech emotion recognition based on SVM using MATLAB
Fernandes et al. Speech emotion recognition using mel frequency cepstral coefficient and SVM classifier
Ahmed et al. CNN-based speech segments endpoints detection framework using short-time signal energy features
Baghel et al. Exploration of excitation source information for shouted and normal speech classification
Ghosal et al. Automatic male-female voice discrimination
WO2020238681A1 (zh) 音频处理方法、装置和人机交互系统
Dalsaniya et al. Development of a novel database in Gujarati language for spoken digits classification
Singh et al. Speaker Recognition Assessment in a Continuous System for Speaker Identification
Akinrinmade et al. Creation of a Nigerian voice corpus for indigenous speaker recognition
Lekshmi et al. An acoustic model and linguistic analysis for Malayalam disyllabic words: a low resource language
Sharma et al. Comparative analysis of various feature extraction techniques for classification of speech disfluencies

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20931026

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 13.02.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20931026

Country of ref document: EP

Kind code of ref document: A1