WO2020192009A1 - 一种基于神经网络的静音检测方法、终端设备及介质 - Google Patents

一种基于神经网络的静音检测方法、终端设备及介质 Download PDF

Info

Publication number
WO2020192009A1
WO2020192009A1 PCT/CN2019/103149 CN2019103149W WO2020192009A1 WO 2020192009 A1 WO2020192009 A1 WO 2020192009A1 CN 2019103149 W CN2019103149 W CN 2019103149W WO 2020192009 A1 WO2020192009 A1 WO 2020192009A1
Authority
WO
WIPO (PCT)
Prior art keywords
audio
layer
subsequence
dimensionality reduction
convolution
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2019/103149
Other languages
English (en)
French (fr)
Inventor
房树明
程宁
王健宗
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
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 Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2020192009A1 publication Critical patent/WO2020192009A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

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
    • G10L25/03—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
    • G10L25/27—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
    • G10L25/78—Detection of presence or absence of voice signals
    • 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
    • G10L2025/783—Detection of presence or absence of voice signals based on threshold decision

Definitions

  • This application belongs to the field of artificial intelligence technology, and in particular relates to a method for detecting silence based on a neural network, a terminal device and a computer non-volatile readable storage medium.
  • Mute detection refers to the feature analysis of audio signals to identify voice signals and noise signals from the audio signals. It has a very wide range of applications in the fields of speech coding, speech enhancement, and speech recognition. Mute detection is the first step in speech coding, speech enhancement and speech recognition. Its accuracy is directly related to whether the subsequent speech processing can be carried out effectively.
  • Traditional silence detection usually uses detection methods such as zero-crossing detection, correlation detection or spectral envelope detection. These detection methods all need to first convert the time domain audio signal into a frequency domain signal, which is not only cumbersome to operate, difficult to apply, but also accurate in detection The rate is low.
  • the embodiments of this application provide a neural network-based silent detection method, terminal equipment, and computer non-volatile readable storage medium to solve the cumbersome operation, difficulty in application, and detection accuracy of the existing silent detection method. Low problem.
  • a silent detection method based on neural network including:
  • the silence detection model is a one-dimensional convolutional neural network model, and the feature value of the audio sub-sequence is used for Characterizing the probability that the audio segment corresponding to the audio subsequence is a speech signal, and the feature value is a one-dimensional value;
  • the feature value of the audio subsequence is greater than or equal to the preset feature value threshold, it is determined that the audio segment corresponding to the audio subsequence is a speech signal.
  • a terminal device including:
  • the first sampling unit is configured to sample an original audio signal to be detected based on a preset sampling frequency to obtain a sampling signal corresponding to the original audio signal;
  • the first audio processing unit is configured to perform frame division processing on the sample signal based on the preset receptive field length to obtain at least two frames of audio subsequences;
  • the feature value calculation unit is configured to input the audio subsequence to a pre-trained silence detection model to obtain the feature value of the audio subsequence;
  • the silence detection model is a one-dimensional convolutional neural network model, and the audio
  • the feature value of the subsequence is used to characterize the probability that the audio segment corresponding to the audio subsequence is a speech signal, and the feature value is a one-dimensional value;
  • the silence detection unit is configured to determine that the audio segment corresponding to the audio subsequence is a voice signal if the feature value of the audio subsequence is greater than or equal to a preset feature value threshold.
  • a terminal device including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor.
  • the processor executes the computer-readable instructions, The steps of the above-mentioned silent detection method based on neural network.
  • a computer non-volatile readable storage medium stores computer readable instructions, which when executed by a processor, realize the above Steps of the silent detection method based on neural network.
  • This application samples the original audio signal based on a preset sampling frequency, and performs frame processing on the sampled signal based on the preset receptive field length to obtain at least a frame audio sub-sequence, and uses a pre-trained silence detection model to The audio subsequence undergoes dimensionality reduction processing, and finally the audio subsequence is converted into a one-dimensional value, and based on the magnitude relationship between the one-dimensional value and the preset feature value threshold, it is determined whether the audio segment corresponding to the audio subsequence is speech signal.
  • the silence detection model when performing silence detection on the original audio signal, there is no need to convert the original audio signal from the time domain to the frequency domain, and only need to be converted into a digital audio signal in the time domain, thereby simplifying the silence detection process.
  • the efficiency of silence detection is improved, and because the silence detection model is obtained by training, various parameters included in the silence detection model can be continuously optimized during the training process, thereby improving the accuracy of silence detection.
  • Fig. 1 is an implementation flowchart of a neural network-based silence detection method provided by an embodiment of the present invention
  • FIG. 2 is a specific implementation flow chart of S3 in a method for silent detection based on neural network provided by an embodiment of the present invention
  • FIG. 3 is an implementation flowchart of a method for detecting silence based on a neural network according to another embodiment of the present invention
  • FIG. 4 is a structural block diagram of a terminal device provided by an embodiment of the present invention.
  • Fig. 5 is a structural block diagram of a terminal device according to another embodiment of the present invention.
  • FIG. 1 is an implementation flowchart of a method for detecting silence based on a neural network according to an embodiment of the present invention.
  • the execution subject of the neural network-based silence detection method is the terminal device.
  • Terminal devices include but are not limited to smartphones, tablets or desktop computers.
  • the silent detection method based on neural network as shown in Figure 1 includes the following steps:
  • S1 Sampling the original audio signal to be detected based on the preset sampling frequency to obtain the sampling signal corresponding to the original audio signal.
  • the original audio signal to be detected is an analog audio signal, which is usually collected by a microphone.
  • the terminal device may sample the original audio signal to be detected based on the preset sampling frequency, and then obtain the sampling signal corresponding to the original audio signal.
  • the sampling signal is a digital audio signal
  • the length of the sampling signal is related to the duration of the original audio signal and the preset sampling frequency
  • the length of the sampling signal is used to identify the number of sampling points included in the sampling signal.
  • the length of the sampling signal N t ⁇ f, where t is the duration of the original audio signal, f is the preset sampling frequency, and N is a positive integer, that is, the duration is t seconds based on the preset sampling frequency f
  • the sampling signal obtained by sampling the original audio signal is an audio sequence of length t ⁇ f.
  • the preset sampling frequency can be set according to actual needs, and there is no restriction here.
  • S2 Perform frame division processing on the sampled signal based on the preset receptive field length to obtain at least two frames of audio subsequences.
  • the preset receptive field length refers to the frame length of a single frame on which the sampled signal is subjected to framing processing, that is, each frame of audio obtained by framing the sampled signal based on the preset receptive field length
  • the length of the sub-sequence is equal to the length of the preset receptive field.
  • the preset receptive field length can be set according to actual needs, and there is no restriction here.
  • the preset receptive field length can be T.
  • the embodiment of the present invention needs to perform the same processing on each frame of audio subsequence obtained by framing processing in the subsequent steps, it is necessary to ensure that the length of each frame of audio subsequence obtained by performing framing processing on the sampled signal is the same as the preset If the length of the receptive field is equal, it is necessary to ensure that the length of the sampled signal is an integer multiple of the preset receptive field length. In actual applications, the length of the sampled signal is not an integer multiple of the preset receptive field length. Therefore, the present invention is implemented In an example, the terminal device also detects whether the length of the sampled signal is an integer multiple of the preset receptive field length before performing framing processing on the sampled signal.
  • the terminal device detects that the length of the sampled signal is not an integer multiple of the preset receptive field length, it adjusts the length of the sampled signal based on the preset length adjustment strategy, so that the length of the sampled signal is the preset Integer multiples of the length of the felt field.
  • the preset length adjustment strategy can be set according to actual needs.
  • the preset length adjustment strategy can be: zero-filling the sampling signal until the length of the sampling signal is an integer multiple of the preset receptive field length.
  • performing zero-filling processing on the sampled signal may specifically be: zero-filling before or after the audio sequence corresponding to the sampled signal.
  • the terminal device After the terminal device adjusts the length of the sampled signal to an integer multiple of the preset receptive field length, it performs frame processing on the adjusted length of the sampled signal based on the preset receptive field length to obtain at least two frames of audio subsequences, each frame of audio subsequence
  • the length of the sequence is T, that is, each frame of audio sub-sequence is composed of T sample values.
  • S3 Input the audio sub-sequence to the pre-trained silence detection model to obtain the feature value of the audio sub-sequence;
  • the silence detection model is a one-dimensional convolutional neural network model, and the feature value of the audio sub-sequence It is used to characterize the probability that the audio segment corresponding to the audio subsequence is a speech signal, and the characteristic value is a one-dimensional value.
  • the silent detection model is based on a preset sample set and a deep learning algorithm is used to train a pre-built one-dimensional convolutional neural network model.
  • the input values and intermediate processing values of the one-dimensional convolutional neural network model described in the embodiment of the present invention are both one-dimensional arrays, and the output value of the one-dimensional convolutional neural network model is a one-dimensional value.
  • Each piece of sample data in the preset sample set is composed of an audio sub-sequence of length T and a feature value corresponding to the audio sub-sequence.
  • the feature value of the audio subsequence is used to characterize the probability that the audio segment corresponding to the audio subsequence is a speech signal. For example, if the audio signal corresponding to a certain audio subsequence is a speech signal, the feature value of the audio subsequence can be set to 1, and if the audio signal corresponding to a certain audio subsequence is a noise signal, the audio subsequence can be The eigenvalue of the sequence is set to 0. It should be noted that the audio segment corresponding to the audio subsequence is used to characterize the audio segment corresponding to the audio subsequence in the original audio signal.
  • the one-dimensional convolutional neural network model includes an input layer, a hidden layer, and an output layer.
  • the input layer includes T input nodes, which are used to respectively receive the T sample values contained in the audio subsequence;
  • the hidden layer is composed of L-layer cascaded dimensionality reduction networks, and each layer of dimensionality reduction network is configured with one for audio
  • the first convolution kernel for dimensionality reduction processing of the subsequence.
  • the first convolution kernel is a one-dimensional array whose length is less than the length of the audio subsequence.
  • the step size of the first convolution kernel may be the same as that of the first convolution kernel.
  • the lengths of the cores are the same; the output layer is equipped with a second convolution core used to reduce the dimensionality of the audio subsequence after the convolution processing of the hidden layer output.
  • the second convolution core is also a one-dimensional array.
  • the length of the convolution kernel is equal to the length of the audio subsequence output by the hidden layer.
  • the audio subsequence of length T contained in each sample data in the preset sample set is used as the input of the one-dimensional convolutional neural network model.
  • the feature value of the audio subsequence contained in each sample data is used as the output of the one-dimensional convolutional neural network model, and then the one-dimensional convolutional neural network model is trained.
  • the terminal device can learn the convolution kernel parameters of the first convolution kernel of each layer of the dimensionality reduction network in the hidden layer of the one-dimensional convolutional neural network model.
  • the output layer of the three-dimensional convolutional neural network model learns the convolution kernel parameters of the second convolution kernel.
  • the convolution kernel parameter refers to the value of each element contained in the convolution kernel.
  • the terminal device after the terminal device performs frame processing on the sampled signal to obtain at least two frames of audio subsequences, all the audio subsequences are input to the pre-trained silence detection model to obtain the feature value of each audio subsequence.
  • S3 can be specifically implemented through S31 to S33 as shown in FIG. 2, which is described in detail as follows:
  • S31 Receive the T sample values included in the audio subsequence through the T input nodes included in the input layer of the silence detection model.
  • the terminal device After the terminal device inputs the audio sub-sequence obtained by the framing process to the pre-trained silence detection model, the T input nodes contained in the input layer of the silence detection model respectively receive the T sample values contained in the audio sub-sequence, and receive The received audio sub-sequence is input to the hidden layer of the silence detection model.
  • the terminal device inputs the audio subsequence received in the input layer of the silence detection model to the hidden layer of the silence detection model, in the hidden layer of the silence detection model, it is based on the first order of each layer of the dimensionality reduction network.
  • the convolution kernel performs convolution processing on the audio sub-sequences received by each layer of the dimensionality reduction network, and obtains the feature array of the audio sub-sequences in the L-th dimensionality reduction network.
  • the terminal device performs convolution processing on the audio sub-sequences received by each layer of the dimensionality reduction network based on the first convolution kernel of each layer of the dimensionality reduction network, which specifically includes the following steps:
  • the first convolution kernel based on the first layer of dimensionality reduction network performs convolution processing on the audio subsequence output by the input layer, and inputs the convolution processed audio subsequence into the second layer of dimensionality reduction network ;
  • the second layer of dimensionality reduction network based on the first convolution kernel of the second layer of dimensionality reduction network, the audio subsequence output by the first layer of dimensionality reduction network will be convolved again after convolution processing.
  • the audio subsequence of is input to the third layer of dimensionality reduction network, and so on, and finally, in the hidden layer of the L-th dimensionality reduction network, based on the first convolution check of the L-th dimensionality reduction network to output the L-1 layer
  • the feature array of the audio sub-sequence is obtained. It should be noted that the length of the audio subsequence feature array is much smaller than the length of the audio subsequence.
  • the length of the audio subsequence input to the silence detection model is determined, in practical applications, the number of layers of the dimensionality reduction network contained in the hidden layer and the number of dimensionality reduction networks in each layer can be determined according to actual needs.
  • the length and step length of the first convolution kernel are flexibly set, so that the length of the feature array finally output by the L-th layer dimensionality reduction network is determined, and the length of the feature array of the audio subsequence can also be used to determine the output layer contains The length and step size of the second convolution kernel.
  • the length of the first convolution kernel is equal to the step size, and the length of the audio subsequence received by each layer of the dimensionality reduction network is an integer multiple of the length of the first convolution kernel of the layer, based on this , S32 may specifically include the following steps:
  • the audio subsequences received by the dimensionality reduction network of each layer are subjected to convolution processing based on the first preset convolution formula in turn;
  • the first preset convolution formula is:
  • Kernel ij is the first convolution kernel of the dimensionality reduction network of the i-th layer
  • the value of the j-th element, k i is the length of Kernel ij
  • Audio (i-1)j is the value of the j-th audio element contained in the audio subsequence output by the dimensionality reduction network of the i-1th layer
  • J + k i of the audio elements of the sequence of audio values dimensionality reduction output from the first network layer comprises i-1, Is the first subsequence contained in the audio subsequence output by the dimensionality reduction network of the i-1th layer
  • the value of each audio element, a i-1 is the length of the audio subsequence output by the dimensionality reduction network of the i-1th layer;
  • the audio subsequence after the convolution processing output by the dimensionality reduction network of the Lth layer is determined as the feature array of the audio subsequence.
  • the audio subsequence received by the layer 1 dimensionality reduction network of the hidden layer is the audio subsequence output by the input layer
  • the audio subsequences received by the layer 2 to layer L dimensionality reduction network of the hidden layer are all It is the audio subsequence output by the upper layer of the dimensionality reduction network after the convolution processing.
  • the audio subsequence to be output is obtained by processing the received audio subsequence based on the first preset convolution formula in the first layer of dimensionality reduction network
  • the terminal device After the terminal device obtains the feature array of the audio subsequence in the L-th dimensionality reduction network of the hidden layer, it inputs the feature array of the audio subsequence to the output layer of the silence detection model.
  • S33 Perform convolution processing on the feature array of the audio subsequence based on the second convolution kernel in the output layer of the silence detection model to obtain the feature value of the audio subsequence.
  • the terminal device performs convolution processing on the feature array of the audio subsequence output by the hidden layer based on the second convolution kernel in the output layer to obtain the feature value of the audio subsequence. It should be noted that, since the length of the second convolution kernel of the output layer is equal to the length of the feature array of the audio subsequence output by the hidden layer, the feature array of the audio subsequence is convolved through the second convolution kernel.
  • the characteristic value of the obtained audio subsequence is a one-dimensional value.
  • S33 can be implemented through the following steps:
  • convolution processing is performed on the feature array of the audio subsequence based on a second preset convolution formula to obtain the feature value of the audio subsequence;
  • the second preset convolution formula is:
  • Audio final is the feature value of the audio subsequence
  • a final is the length of the feature array of the audio subsequence
  • Kernel j is the value of the jth element in the second convolution kernel
  • Audio j is the value The value of the j-th audio element in the feature array of the audio subsequence.
  • the terminal device After the terminal device calculates the feature value of each audio subsequence, it compares the feature value of each audio subsequence with a preset feature value threshold. If the terminal device detects that the characteristic value of a certain audio subsequence is greater than or equal to the preset characteristic value threshold, it determines that the audio segment corresponding to the audio subsequence is a speech signal.
  • the preset feature value threshold can be set according to actual needs, and there is no limitation here.
  • the neural network-based silence detection method may further include S5.
  • S5 the neural network-based silence detection method
  • the terminal device detects that the characteristic value of a certain audio subsequence is less than the preset characteristic value threshold, it determines that the audio segment corresponding to the audio subsequence is a noise signal.
  • the embodiment of the present invention samples the original audio signal based on the preset sampling frequency, and performs framing processing on the sampled sample signal based on the preset receptive field length to obtain at least a frame audio sub-sequence.
  • the trained silence detection model performs dimensionality reduction processing on the audio subsequence, and finally converts the audio subsequence into a one-dimensional value, and determines the audio subsequence based on the magnitude relationship between the one-dimensional value and the preset feature value threshold Whether the corresponding audio segment is a voice signal.
  • the silence detection model when performing silence detection on the original audio signal, there is no need to convert the original audio signal from the time domain to the frequency domain, and only need to be converted into a digital audio signal in the time domain, thereby simplifying the silence detection process.
  • the efficiency of silence detection is improved, and because the silence detection model is obtained by training, various parameters included in the silence detection model can be continuously optimized during the training process, thereby improving the accuracy of silence detection.
  • FIG. 4 is a structural block diagram of a terminal device according to an embodiment of the present invention.
  • the terminal device in this embodiment may be a terminal device such as a smart phone or a tablet computer.
  • the units included in the terminal device are used to execute the steps in the embodiments corresponding to FIGS. 1 to 3.
  • the terminal device 400 includes: a first sampling unit 41, a first audio processing unit 42, a feature value calculation unit 43, and a silence detection unit 44. among them:
  • the first sampling unit 41 is configured to sample the original audio signal to be detected based on a preset sampling frequency to obtain a sampling signal corresponding to the original audio signal;
  • the first audio processing unit 42 is configured to perform frame division processing on the sampled signal based on the preset receptive field length to obtain at least two frames of audio subsequences;
  • the feature value calculation unit 43 is configured to input the audio subsequence into a pre-trained silence detection model to obtain the feature value of the audio subsequence;
  • the silence detection model is a one-dimensional convolutional neural network model, and
  • the feature value of the audio subsequence is used to characterize the probability that the audio segment corresponding to the audio subsequence is a speech signal, and the feature value is a one-dimensional value;
  • the silence detection unit 44 is configured to determine that the audio segment corresponding to the audio subsequence is a voice signal if the feature value of the audio subsequence is greater than or equal to a preset feature value threshold.
  • each frame of the audio subsequence includes T sample values;
  • the silence detection model includes an input layer, a hidden layer, and an output layer.
  • the input layer includes T input nodes, and the hidden layer consists of
  • the L-layer cascaded dimensionality reduction network constitutes, each layer of the dimensionality reduction network is configured with a first convolution kernel;
  • the feature value calculation unit 43 specifically includes: a first receiving unit, a first calculation unit, and a second calculation unit. among them:
  • the first receiving unit is configured to receive the T sample values included in the audio subsequence through the T input nodes included in the input layer of the silence detection model.
  • the first calculation unit is configured to perform convolution processing on the audio subsequences received by each layer of the dimensionality reduction network based on the first convolution kernel of each layer of the dimensionality reduction network in the hidden layer of the silence detection model, Obtain the feature array of the audio subsequence in the dimensionality reduction network of the Lth layer.
  • the second calculation unit is configured to perform convolution processing on the feature array of the audio subsequence based on the second convolution kernel in the output layer of the silence detection model to obtain the feature value of the audio subsequence.
  • the length of the first convolution kernel is equal to its step size, and the length of the audio subsequence received by the dimensionality reduction network of each layer is an integer multiple of the length of the first convolution kernel of the layer ;
  • the first calculation unit is specifically configured to:
  • the audio subsequences received by the dimensionality reduction network of each layer are subjected to convolution processing based on the first preset convolution formula in turn;
  • the first preset convolution formula is:
  • the audio subsequences received by the dimensionality reduction network of each layer are subjected to convolution processing based on the first preset convolution formula in turn;
  • the first preset convolution formula is:
  • Kernel ij is the first convolution kernel of the dimensionality reduction network of the i-th layer
  • the value of the j-th element, k i is the length of Kernel ij
  • Audio (i-1)j is the value of the j-th audio element contained in the audio subsequence output by the dimensionality reduction network of the i-1th layer
  • J + k i of the audio elements of the sequence of audio values dimensionality reduction output from the first network layer comprises i-1, Is the first audio subsequence included in the output of the dimensionality reduction network of the i-1th layer
  • the value of each audio element, a i-1 is the length of the audio subsequence output by the dimensionality reduction network of the i-1th layer;
  • the audio subsequence after the convolution processing output by the dimensionality reduction network of the Lth layer is determined as the feature array of the audio subsequence.
  • convolution processing is performed on the feature array of the audio subsequence based on a second preset convolution formula to obtain the feature value of the audio subsequence;
  • the second preset convolution formula is:
  • Audio final is the feature value of the audio subsequence
  • a final is the length of the feature array of the audio subsequence
  • Kernel j is the value of the jth element in the second convolution kernel
  • Audio j is the value The value of the j-th audio element in the feature array of the audio subsequence.
  • the silence detection unit 44 is further configured to determine that the audio segment corresponding to the audio subsequence is a noise signal if the feature value of the audio subsequence is less than a preset feature value threshold.
  • the terminal device samples the original audio signal based on the preset sampling frequency, and performs frame processing on the sampled signal based on the preset receptive field length to obtain at least a frame audio subsequence.
  • the audio sub-sequence is finally converted into a one-dimensional value, and based on the magnitude relationship between the one-dimensional value and the preset feature value threshold, the Whether the audio segment corresponding to the audio subsequence is a speech signal.
  • the silence detection model when performing silence detection on the original audio signal, there is no need to convert the original audio signal from the time domain to the frequency domain, and only need to be converted into a digital audio signal in the time domain, thereby simplifying the silence detection process.
  • the efficiency of silence detection is improved, and because the silence detection model is obtained by training, various parameters included in the silence detection model can be continuously optimized during the training process, thereby improving the accuracy of silence detection.
  • Fig. 5 is a structural block diagram of a terminal device according to another embodiment of the present invention.
  • the terminal device 5 of this embodiment includes: a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and running on the processor 50, such as a neural network-based mute Procedures for testing methods.
  • the processor 50 executes the computer program 52, the steps in each embodiment of the above-mentioned neural network-based silence detection method are implemented, for example, S1 to S4 shown in FIG. 1.
  • the processor 50 executes the computer program 52
  • the functions of the units in the embodiment corresponding to FIG. 4 are realized, for example, the functions of the units 41 to 44 shown in FIG. 4, please refer to the corresponding implementation in FIG. 4 for details The relevant description in the example will not be repeated here.
  • the computer program 52 may be divided into one or more units, and the one or more units are stored in the memory 51 and executed by the processor 50 to complete the present invention.
  • the one or more units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program 52 in the terminal device 5.
  • the computer program 52 may be divided into a first sampling unit, a first audio processing unit, a feature value calculation unit, and a silence detection unit, and the specific functions of each unit are as described above.
  • the terminal device may include, but is not limited to, a processor 50 and a memory 51.
  • FIG. 5 is only an example of the terminal device 5, and does not constitute a limitation on the terminal device 5. It may include more or less components than shown in the figure, or a combination of certain components, or different components.
  • the terminal device may also include input and output devices, network access devices, buses, etc.
  • the so-called processor 50 may be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory 51 may be an internal storage unit of the terminal device 5, such as a hard disk or a memory of the terminal device 5.
  • the memory 51 may also be an external storage device of the terminal device 5, such as a plug-in hard disk equipped on the terminal device 5, a smart memory card (Smart Media Card, SMC), or a Secure Digital (SD). Card, Flash Card, etc. Further, the memory 51 may also include both an internal storage unit of the terminal device 5 and an external storage device.
  • the memory 51 is used to store the computer program and other programs and data required by the terminal device.
  • the memory 51 can also be used to temporarily store data that has been output or will be output.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • the integrated module/unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • this application implements all or part of the procedures in the above-mentioned embodiments and methods, and can also be completed by instructing relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. in.
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • ROM read only memory
  • PROM programmable ROM
  • EPROM electrically programmable ROM
  • EEPROM electrically erasable programmable ROM
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Circuit For Audible Band Transducer (AREA)

Abstract

人工智能技术领域的一种基于神经网络的静音检测方法、终端设备及计算机非易失性可读存储介质，通过基于预设采样频率对待检测的原始音频信号进行采样，得到原始音频信号对应的采样信号（S1）；基于预设感受野长度对采样信号进行分帧处理，得到至少两帧音频子序列（S2）；将音频子序列输入至预先训练好的静音检测模型，得到音频子序列的特征值；静音检测模型为一维卷积神经网络模型，音频子序列的特征值用于表征音频子序列对应的音频段为语音信号的概率，特征值为一维数值（S3）；若音频子序列的特征值大于或等于预设特征值阈值，则确定音频子序列对应的音频段为语音信号（S4），从而提高了静音检测的效率及准确率。

Description

一种基于神经网络的静音检测方法、终端设备及介质
本申请申明享有2019年3月25日递交的申请号为201910226470.2、名称为“一种基于神经网络的静音检测方法、终端设备及介质”中国专利申请的优先权，该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本申请属于人工智能技术领域，尤其涉及一种基于神经网络的静音检测方法、终端设备及计算机非易失性可读存储介质。
背景技术
静音检测是指对音频信号进行特征分析，从音频信号中识别出语音信号和噪声信号，它在语音编码、语音增强及语音识别等领域具有非常广泛的应用。静音检测作为语音编码、语音增强及语音识别等的第一步，其准确率直接关系到后续的语音处理工作能否有效进行。
传统的静音检测通常采用过零检测、相关性检测或谱包络检测等检测方法，这些检测方法均需要先将时域音频信号转换为频域信号，不仅操作繁琐，应用难度大，而且检测准确率较低。
技术问题
本申请实施例提供了一种基于神经网络的静音检测方法、终端设备及计算机非易失性可读存储介质，以解决现有的静音检测方法存在的操作繁琐，应用难度较大，检测准确率低的问题。
技术解决方案
第一方面，提供了一种基于神经网络的静音检测方法，包括：
基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对应的采样信号；
基于预设感受野长度对所述采样信号进行分帧处理，得到至少两帧音频子序列；
将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值；
若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应的音频段为语音信号。
第二方面，提供了一种终端设备，包括：
第一采样单元，用于基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对应的采样信号；
第一音频处理单元，用于基于预设感受野长度对所述采样信号进行分帧处理，得到至少两帧音频子序列；
特征值计算单元，用于将所述音频子序列输入至预先训练好的静音检测模型，得到所 述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值；
静音检测单元，用于若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应的音频段为语音信号。
第三方面，提供了一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，所述处理器执行所述计算机可读指令时上述的基于神经网络的静音检测方法的步骤。
第四方面，提供了一种计算机非易失性可读存储介质，所述计算机非易失性可读存储介质存储有计算机可读指令，所述计算机可读指令被处理器执行时实现上述的基于神经网络的静音检测方法的步骤。
有益效果
本申请基于预设采样频率对原始音频信号进行采样，并基于预设感受野长度对采样得到的采样信号进行分帧处理，得到至少量帧音频子序列，通过采用预先训练好的静音检测模型对音频子序列进行降维处理，最终将音频子序列转换为一个一维数值，并基于该一维数值与预设特征值阈值之间的大小关系，确定该音频子序列对应的音频段是否为语音信号。本发明实施例在对原始音频信号进行静音检测时，无需将原始音频信号进行时域到频域的转换，只需在时域将其转换为数字音频信号即可，从而简化了静音检测流程，提高了静音检测的效率，且由于静音检测模型是训练得到的，因而在训练过程中可以不断对静音检测模型包含的各项参数进行优化，从而能够提高静音检测的准确率。
附图说明
图1是本发明实施例提供的一种基于神经网络的静音检测方法的实现流程图；
图2是本发明实施例提供的一种基于神经网络的静音检测方法中S3的具体实现流程图；
图3是本发明另一实施例提供的一种基于神经网络的静音检测方法的实现流程图；
图4是本发明实施例提供的一种终端设备的结构框图；
图5是本发明另一实施例提供的一种终端设备的结构框图。
本发明的实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
请参阅图1，图1是本发明实施例提供的一种基于神经网络的静音检测方法的实现流程图。本实施例中，基于神经网络的静音检测方法的执行主体为终端设备。终端设备包括但不限于智能手机、平板电脑或台式电脑。如图1所示的基于神经网络的静音检测方法包括以下步骤：
S1：基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对 应的采样信号。
待检测的原始音频信号为模拟音频信号，其通常通过麦克风采集得到。在对原始音频信号进行静音检测时，需要先将原始音频信号转换为数字音频信号。本发明实施例中，终端设备可以基于预设采样频率对待检测的原始音频信号进行采样，进而得到原始音频信号对应的采样信号。其中，采样信号为数字音频信号，采样信号的长度与原始音频信号的时长及预设采样频率有关，采样信号的长度用于标识采样信号所包含的采样点的个数。具体的，采样信号的长度N＝t×f，其中，t为原始音频信号的时长，f为预设采样频率，N为正整数，也就是说，基于预设采样频率f对时长为t秒的原始音频信号进行采样得到的采样信号是一个长度为t×f的音频序列。在实际应用中，预设采样频率可以根据实际需求设置，此处不做限制。
S2：基于预设感受野长度对所述采样信号进行分帧处理，得到至少两帧音频子序列。
本发明实施例中，预设感受野长度指对采样信号进行分帧处理时所依据的单帧帧长，也就是说，基于预设感受野长度对采样信号进行分帧处理得到的每帧音频子序列的长度与预设感受野长度相等。预设感受野长度可根据实际需求设置，此处不做限制，例如，预设感受野长度可以为T。
由于本发明实施例在后续步骤中需要对分帧处理得到的每帧音频子序列均进行相同的处理，因此需保证对采样信号进行分帧处理得到的每帧音频子序列的长度均与预设感受野长度相等，这就需要保证采样信号的长度是预设感受野长度的整数倍，而实际应用中，采样信号的长度并不都是预设感受野长度的整数倍，因此，本发明实施例中，终端设备在对采样信号进行分帧处理之前，还检测采样信号的长度是否是预设感受野长度的整数倍。
作为本发明一实施例，终端设备若检测到采样信号的长度不是预设感受野长度的整数倍，则基于预设长度调整策略对采样信号的长度进行调整，以使采样信号的长度为预设感受野长度的整数倍。其中，预设长度调整策略可以根据实际需求设置，示例性的，预设长度调整策略可以是：对采样信号进行补0处理，直至采样信号的长度为预设感受野长度的整数倍。实际应用中，对采样信号进行补0处理具体可以为：在采样信号对应的音频序列之前或之后补0。
终端设备将采样信号的长度调整为预设感受野长度的整数倍后，基于预设感受野长度对调整了长度后的采样信号进行分帧处理，得到至少两帧音频子序列，每帧音频子序列的长度为T，即每帧音频子序列由T个采样值构成。本发明实施例中，假设对采样信号进行长度调整时在采样信号对应的音频序列中添加了n个0，则对调整了长度后的采样信号进行分帧处理后得到的音频子序列的个数m＝(t×f+n)/T。
S3：将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值。
静音检测模型是基于预设样本集，采用深度学习算法对预先构建的一维卷积神经网络模型进行训练得到的。需要说明的是，本发明实施例所述的一维卷积神经网络模型的输入值及中间处理值均为一维数组，一维卷积神经网络模型的输出值为一维数值。
预设样本集中的每条样本数据均由长度为T的音频子序列及该音频子序列对应的特征值构成。音频子序列的特征值用于表征该音频子序列对应的音频段为语音信号的概率。例如，若某一音频子序列对应的音频信号为语音信号，则可以将该音频子序列的特征值设为1，若某一音频子序列对应的音频信号为噪声信号，则可以将该音频子序列的特征值设为0。需要说明的是，音频子序列对应的音频段用于表征音频子序列在原始音频信号中所对应的音频片段。
本发明实施例中，一维卷积神经网络模型包括输入层、隐藏层及输出层。其中，输入层包括T个输入节点，用于分别接收音频子序列所包含的T个采样值；隐藏层由L层级联的降维网络构成，每层降维网络均配置有一个用于对音频子序列进行降维处理的第一卷积核，第一卷积核为一维数组，其长度小于音频子序列的长度，本实施例中，第一卷积核的步长可以与第一卷积核的长度相等；输出层配置有一个用于对隐藏层输出的卷积处理后的音频子序列进行降维处理的第二卷积核，第二卷积核也为一维数组，第二卷积核的长度与隐藏层输出的音频子序列的长度相等。
在基于预设样本集对预先构建的一维卷积神经网络模型进行训练时，将预设样本集中每条样本数据包含的长度为T的音频子序列作为一维卷积神经网络模型的输入，将每条样本数据包含的音频子序列的特征值作为一维卷积神经网络模型的输出，进而对一维卷积神经网络模型进行训练。在对一维卷积神经网络模型进行训练的过程中，终端设备可以在一维卷积神经网络模型的隐藏层学习到各层降维网络的第一卷积核的卷积核参数，在一维卷积神经网络模型的输出层学习到第二卷积核的卷积核参数。卷积核参数指卷积核所包含的各个元素的值。在完成对一维卷积神经网络模型的训练后，终端设备将训练完成的一维卷积神经网络模型确定为静音检测模型，也就是说，本发明实施例中的静音检测模型包括输入层、隐藏层及输出层。
本发明实施例中，终端设备对采样信号进行分帧处理得到至少两帧音频子序列后，将所有音频子序列均输入至预先训练好的静音检测模型，得到每一音频子序列的特征值。
具体的，作为本发明一实施例，S3具体可以通过如图2所示的S31～S33来实现，详述如下：
S31：通过所述静音检测模型的输入层包含的T个输入节点分别接收所述音频子序列包含的T个采样值。
终端设备将分帧处理得到的音频子序列输入至预先训练好的静音检测模型后，通过静音检测模型的输入层包含的T个输入节点分别接收音频子序列包含的T个采样值，并将接收到的音频子序列输入至静音检测模型的隐藏层。
S32：在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组。
本实施例中，终端设备将在静音检测模型的输入层接收到的音频子序列输入至静音检测模型的隐藏层后，在静音检测模型的隐藏层中，依次基于每层降维网络的第一卷积核对每层降维网络接收到的音频子序列进行卷积处理，在第L层降维网络中得到音频子序列的特征数组。
终端设备在静音检测模型的隐藏层中，依次基于每层降维网络的第一卷积核对每层降维网络接收到的音频子序列进行卷积处理，具体包括以下步骤：在隐藏层包含的第1层降维网络中基于第1层降维网络的第一卷积核对输入层输出的音频子序列进行卷积处理，将卷积处理后的音频子序列输入至第2层降维网络中；在第2层降维网络中基于第2层降维网络的第一卷积核对第1层降维网络输出的卷积处理后的音频子序列再次进行卷积处理，将再次卷积处理后的音频子序列输入至第3层降维网络中，以此类推，最终，在隐藏层的第L层降维网络中基于第L层降维网络的第一卷积核对第L-1层输出的卷积处理后的音频子序列进行卷积处理后，即得到音频子序列的特征数组。需要说明的是，音频子序列特征数组的长度远小于音频子序列的长度。
需要说明的是，由于输入至静音检测模型的音频子序列的长度是确定的，因此，在实际应用中，可根据实际需求对隐藏层所包含的降维网络的层数、每层降维网络的第一卷积核的长度及步长进行灵活设置，使得第L层降维网络最终输出的特征数组的长度是确定的，进而也可基于音频子序列的特征数组的长度确定输出层所包含的第二卷积核的长度及步长。
作为本发明一实施例，第一卷积核的长度与步长相等，且每层降维网络接收到的音频子序列的长度为该层的第一卷积核的长度的整数倍，基于此，S32具体可以包括以下步骤：
在所述静音检测模型的隐藏层中，依次基于第一预设卷积公式对每层所述降维网络接收到的音频子序列进行卷积处理；所述第一预设卷积公式为：
Figure PCTCN2019103149-appb-000001
其中，Audio i为第i层所述降维网络待输出的音频子序列，i＝{1,2,…,L}，Kernel ij为第 i层所述降维网络的第一卷积核中第j个元素的值，k i为Kernel ij的长度，Audio (i-1)j为第i-1层所述降维网络输出的音频子序列包含的第j个音频元素的值，
Figure PCTCN2019103149-appb-000002
为第i-1层所述降维网络输出的音频子序列包含的第j+k i个音频元素的值，
Figure PCTCN2019103149-appb-000003
为第i-1层所述降维网络输出的音频子序列包含的第
Figure PCTCN2019103149-appb-000004
个音频元素的值，a i-1为第i-1层所述降维网络输出的音频子序列的长度；
将第L层所述降维网络输出的卷积处理后的所述音频子序列确定为所述音频子序列的特征数组。
本发明实施例中，隐藏层的第1层降维网络接收到的音频子序列为输入层输出的音频子序列，隐藏层的第2层至第L层降维网络接收到的音频子序列均为其上一层降维网络输出的进行卷积处理后的音频子序列。
示例性的，假设隐藏层的第1层降维网络接收到的音频子序列为Audio 0＝[0.1,0.7,-0.3,0.9,0.6,-0.4,0.85,0.5]，第1层降维网络的第一卷积核Kernel 1＝[0.6,0.8]，则在第1层降维网络中基于第一预设卷积公式对接收到的音频子序列进行处理得到的待输出的音频子序列
Figure PCTCN2019103149-appb-000005
终端设备在隐藏层的第L层降维网络中得到音频子序列的特征数组后，将音频子序列的特征数组输入至静音检测模型的输出层。
S33：在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值。
本实施例中，终端设备在输出层中基于第二卷积核对隐藏层输出的音频子序列的特征数组进行卷积处理，得到音频子序列的特征值。需要说明的是，由于输出层的第二卷积核的长度与隐藏层输出的音频子序列的特征数组的长度相等，因此，通过第二卷积核对音频子序列的特征数组进行卷积处理后得到的音频子序列的特征值为一个一维数值。
作为本发明一实施例，S33可以通过以下步骤实现：
在所述输出层中基于第二预设卷积公式对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值；所述第二预设卷积公式为：
Figure PCTCN2019103149-appb-000006
其中，Audio final为所述音频子序列的特征值，a final为所述音频子序列的特征数组的长 度，Kernel j为所述第二卷积核中第j个元素的值，Audio j为所述音频子序列的特征数组中第j个音频元素的值。
示例性的，假设终端设备在隐藏层的第L层降维网络中得到的音频子序列的特征数组Audio＝[-0.47,-0.38]，第二卷积核Kernel j＝[-0.9,0.2]，则在输出层基于第二预设卷积公式对音频子序列的特征数组进行卷积处理得到的音频子序列的特征值Audio final＝-0.9×(-0.47)+0.2×(-0.38)＝0.35。
S4：若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应的音频段为语音信号。
本发明实施例中，终端设备计算得到每一音频子序列的特征值后，将每一音频子序列的特征值与预设特征值阈值进行比较。终端设备若检测到某一音频子序列的特征值大于或等于预设特征值阈值，则确定该音频子序列对应的音频段为语音信号。其中，预设特征值阈值可根据实际需求设置，此处不做限制。
作为本发明另一实施例，如图3所示，在S3之后，基于神经网络的静音检测方法还可以包括S5。详述如下：
S5：若所述音频子序列的特征值小于预设特征值阈值，则确定所述音频子序列对应的音频段为噪声信号。
本实施例中，终端设备若检测到某一音频子序列的特征值小于预设特征值阈值，则确定该音频子序列对应的音频段为噪声信号。
以上可以看出，本发明实施例基于预设采样频率对原始音频信号进行采样，并基于预设感受野长度对采样得到的采样信号进行分帧处理，得到至少量帧音频子序列，通过采用预先训练好的静音检测模型对音频子序列进行降维处理，最终将音频子序列转换为一个一维数值，并基于该一维数值与预设特征值阈值之间的大小关系，确定该音频子序列对应的音频段是否为语音信号。本发明实施例在对原始音频信号进行静音检测时，无需将原始音频信号进行时域到频域的转换，只需在时域将其转换为数字音频信号即可，从而简化了静音检测流程，提高了静音检测的效率，且由于静音检测模型是训练得到的，因而在训练过程中可以不断对静音检测模型包含的各项参数进行优化，从而能够提高静音检测的准确率。
请参阅图4，图4是本发明实施例提供的一种终端设备的结构框图。本实施例中的终端设备可以是智能手机、平板电脑等终端设备。该终端设备包括的各单元用于执行图1至图3对应的实施例中的各步骤。具体请参阅图1至图3以及图1至图3所对应的实施例中的相关描述。为了便于说明，仅示出了与本实施例相关的部分。参见图4，终端设备400包括：第一采样单元41、第一音频处理单元42、特征值计算单元43及静音检测单元44。其中：
第一采样单元41，用于基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对应的采样信号；
第一音频处理单元42，用于基于预设感受野长度对所述采样信号进行分帧处理，得到 至少两帧音频子序列；
特征值计算单元43，用于将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值；
静音检测单元44，用于若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应的音频段为语音信号。
作为本发明一实施例，每帧所述音频子序列包含T个采样值；所述静音检测模型包括输入层、隐藏层及输出层，所述输入层包含T个输入节点，所述隐藏层由L层级联的降维网络构成，每层所述降维网络均配置有第一卷积核；特征值计算单元43具体包括：第一接收单元、第一计算单元及第二计算单元。其中：
第一接收单元用于通过所述静音检测模型的输入层包含的T个输入节点分别接收所述音频子序列包含的T个采样值。
第一计算单元用于在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组。
第二计算单元用于在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值。
作为本发明一实施例，所述第一卷积核的长度与其步长相等，每层所述降维网络接收到的音频子序列的长度为该层的第一卷积核的长度的整数倍；所述第一计算单元具体用于：
在所述静音检测模型的隐藏层中，依次基于第一预设卷积公式对每层所述降维网络接收到的音频子序列进行卷积处理；所述第一预设卷积公式为：
在所述静音检测模型的隐藏层中，依次基于第一预设卷积公式对每层所述降维网络接收到的音频子序列进行卷积处理；所述第一预设卷积公式为：
Figure PCTCN2019103149-appb-000007
其中，Audio i为第i层所述降维网络待输出的音频子序列，i＝{1,2,…,L}，Kernel ij为第 i层所述降维网络的第一卷积核中第j个元素的值，k i为Kernel ij的长度，Audio (i-1)j为第i-1层所述降维网络输出的音频子序列包含的第j个音频元素的值，
Figure PCTCN2019103149-appb-000008
为第i-1层所述降维网络输出的音频子序列包含的第j+k i个音频元素的值，
Figure PCTCN2019103149-appb-000009
为第i-1层所述降维网络输出的音频子序列包含的第
Figure PCTCN2019103149-appb-000010
个音频元素的值，a i-1为第i-1层所述降维网络输出的音频子序列的长度；
将第L层所述降维网络输出的卷积处理后的所述音频子序列确定为所述音频子序列的特征数组。
作为本发明一实施例，所述第二卷积核的长度与所述音频子序列的特征数组的长度相等；所述第二计算单元具体用于：
在所述输出层中基于第二预设卷积公式对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值；所述第二预设卷积公式为：
Figure PCTCN2019103149-appb-000011
其中，Audio final为所述音频子序列的特征值，a final为所述音频子序列的特征数组的长度，Kernel j为所述第二卷积核中第j个元素的值，Audio j为所述音频子序列的特征数组中第j个音频元素的值。
作为本发明一实施例，静音检测单元44还用于若所述音频子序列的特征值小于预设特征值阈值，则确定所述音频子序列对应的音频段为噪声信号。
以上可以看出，本实施例提供的终端设备基于预设采样频率对原始音频信号进行采样，并基于预设感受野长度对采样得到的采样信号进行分帧处理，得到至少量帧音频子序列，通过采用预先训练好的静音检测模型对音频子序列进行降维处理，最终将音频子序列转换为一个一维数值，并基于该一维数值与预设特征值阈值之间的大小关系，确定该音频子序列对应的音频段是否为语音信号。本发明实施例在对原始音频信号进行静音检测时，无需将原始音频信号进行时域到频域的转换，只需在时域将其转换为数字音频信号即可，从而简化了静音检测流程，提高了静音检测的效率，且由于静音检测模型是训练得到的，因而在训练过程中可以不断对静音检测模型包含的各项参数进行优化，从而能够提高静音检测的准确率。
图5是本发明另一实施例提供的一种终端设备的结构框图。如图5所示，该实施例的终端设备5包括：处理器50、存储器51以及存储在所述存储器51中并可在所述处理器50上运行的计算机程序52，例如基于神经网络的静音检测方法的程序。处理器50执行所述计算机程序52时实现上述各个基于神经网络的静音检测方法各实施例中的步骤，例如图1所 示的S1至S4。或者，所述处理器50执行所述计算机程序52时实现上述图4对应的实施例中各单元的功能，例如，图4所示的单元41至44的功能，具体请参阅图4对应的实施例中的相关描述，此处不赘述。
示例性的，所述计算机程序52可以被分割成一个或多个单元，所述一个或者多个单元被存储在所述存储器51中，并由所述处理器50执行，以完成本发明。所述一个或多个单元可以是能够完成特定功能的一系列计算机程序指令段，该指令段用于描述所述计算机程序52在所述终端设备5中的执行过程。例如，所述计算机程序52可以被分割成第一采样单元、第一音频处理单元、特征值计算单元及静音检测单元，各单元具体功能如上所述。
所述终端设备可包括，但不仅限于，处理器50、存储器51。本领域技术人员可以理解，图5仅仅是终端设备5的示例，并不构成对终端设备5的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如所述终端设备还可以包括输入输出设备、网络接入设备、总线等。
所称处理器50可以是中央处理单元(Central Processing Unit，CPU)，还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器51可以是所述终端设备5的内部存储单元，例如终端设备5的硬盘或内存。所述存储器51也可以是所述终端设备5的外部存储设备，例如所述终端设备5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。进一步地，所述存储器51还可以既包括所述终端设备5的内部存储单元也包括外部存储设备。所述存储器51用于存储所述计算机程序以及所述终端设备所需的其他程序和数据。所述存储器51还可以用于暂时地存储已经输出或者将要输出的数据。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中，上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。另外，各功能单元、模块的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。上述系统中单元、模块的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述或记载的部分，可以参见其它实施例的相关描述。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
所述集成的模块/单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实现上述实施例方法中的全部或部分流程，也可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一计算机可读存储介质中。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种基于神经网络的静音检测方法，其特征在于，包括：
    基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对应的采样信号；
    基于预设感受野长度对所述采样信号进行分帧处理，得到至少两帧音频子序列；
    将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值；
    若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应的音频段为语音信号。
  2. 根据权利要求1所述的基于神经网络的静音检测方法，其特征在于，每帧所述音频子序列包含T个采样值；所述静音检测模型包括输入层、隐藏层及输出层，所述输入层包含T个输入节点，所述隐藏层由L层级联的降维网络构成，每层所述降维网络均配置有第一卷积核；
    所述将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值，包括：
    通过所述静音检测模型的输入层包含的T个输入节点分别接收所述音频子序列包含的T个采样值；
    在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组；
    在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值。
  3. 根据权利要求2所述的基于神经网络的静音检测方法，其特征在于，所述第一卷积核的长度与其步长相等，每层所述降维网络接收到的音频子序列的长度为该层的第一卷积核的长度的整数倍；
    所述在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组，包括：
    在所述静音检测模型的隐藏层中，依次基于第一预设卷积公式对每层所述降维网络接收到的音频子序列进行卷积处理；所述第一预设卷积公式为：
    Figure PCTCN2019103149-appb-100001
    其中，Audio i为第i层所述降维网络待输出的音频子序列，i＝{1,2,…,L}，Kernel ij为第i层所述降维网络的第一卷积核中第j个元素的值，k i为Kernel ij的长度，Audio (i-1)j为第i-1层所述降维网络输出的音频子序列包含的第j个音频元素的值，
    Figure PCTCN2019103149-appb-100002
    为第i-1层所述降维网络输出的音频子序列包含的第j+k i个音频元素的值，
    Figure PCTCN2019103149-appb-100003
    为第i-1层所述降维网络输出的音频子序列包含的第
    Figure PCTCN2019103149-appb-100004
    个音频元素的值，a i-1为第i-1层所述降维网络输出的音频子序列的长度；
    将第L层所述降维网络输出的卷积处理后的所述音频子序列确定为所述音频子序列的特征数组。
  4. 根据权利要求2所述的基于神经网络的静音检测方法，其特征在于，所述第二卷积核的长度与所述音频子序列的特征数组的长度相等；
    所述在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值，包括：
    在所述输出层中基于第二预设卷积公式对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值；所述第二预设卷积公式为：
    Figure PCTCN2019103149-appb-100005
    其中，Audio final为所述音频子序列的特征值，a final为所述音频子序列的特征数组的长度，Kernel j为所述第二卷积核中第j个元素的值，Audio j为所述音频子序列的特征数组中第j个音频元素的值。
  5. 根据权利要求1-4任一项所述的基于神经网络的静音检测方法，其特征在于，所述将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值之后，还包括：
    若所述音频子序列的特征值小于预设特征值阈值，则确定所述音频子序列对应的音频 段为噪声信号。
  6. 一种终端设备，其特征在于，包括：
    第一采样单元，用于基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对应的采样信号；
    第一音频处理单元，用于基于预设感受野长度对所述采样信号进行分帧处理，得到至少两帧音频子序列；
    特征值计算单元，用于将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值；
    静音检测单元，用于若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应的音频段为语音信号。
  7. 根据权利要求6所述的终端设备，其特征在于，每帧所述音频子序列包含T个采样值；所述静音检测模型包括输入层、隐藏层及输出层，所述输入层包含T个输入节点，所述隐藏层由L层级联的降维网络构成，每层所述降维网络均配置有第一卷积核；所述终端设备还包括：
    第一接收单元，用于通过所述静音检测模型的输入层包含的T个输入节点分别接收所述音频子序列包含的T个采样值；
    第一计算单元，用于在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组；
    第二计算单元，用于在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值。
  8. 根据权利要求7所述的终端设备，其特征在于，所述第一卷积核的长度与其步长相等，每层所述降维网络接收到的音频子序列的长度为该层的第一卷积核的长度的整数倍；所述第一计算单元具体用于：
    在所述静音检测模型的隐藏层中，依次基于第一预设卷积公式对每层所述降维网络接收到的音频子序列进行卷积处理；所述第一预设卷积公式为：
    Figure PCTCN2019103149-appb-100006
    其中，Audio i为第i层所述降维网络待输出的音频子序列，i＝{1,2,…,L}，Kernel ij为第i层所述降维网络的第一卷积核中第j个元素的值，k i为Kernel ij的长度，Audio (i-1)j为第i-1层所述降维网络输出的音频子序列包含的第j个音频元素的值，
    Figure PCTCN2019103149-appb-100007
    为第i-1层所述降维网络输出的音频子序列包含的第j+k i个音频元素的值，
    Figure PCTCN2019103149-appb-100008
    为第i-1层所述降维网络输出的音频子序列包含的第
    Figure PCTCN2019103149-appb-100009
    个音频元素的值，a i-1为第i-1层所述降维网络输出的音频子序列的长度；
    将第L层所述降维网络输出的卷积处理后的所述音频子序列确定为所述音频子序列的特征数组。
  9. 根据权利要求7所述的终端设备，其特征在于，所述第二卷积核的长度与所述音频子序列的特征数组的长度相等；所述第二计算单元具体用于：
    在所述输出层中基于第二预设卷积公式对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值；所述第二预设卷积公式为：
    Figure PCTCN2019103149-appb-100010
    其中，Audio final为所述音频子序列的特征值，a final为所述音频子序列的特征数组的长度，Kernel j为所述第二卷积核中第j个元素的值，Audio j为所述音频子序列的特征数组中第j个音频元素的值。
  10. 根据权利要求6-9任一项所述的终端设备，其特征在于，所述静音检测单元还用于：若所述音频子序列的特征值小于预设特征值阈值，则确定所述音频子序列对应的音频段为噪声信号。
  11. 一种终端设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现如下步骤：
    基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对应的采样信号；
    基于预设感受野长度对所述采样信号进行分帧处理，得到至少两帧音频子序列；
    将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值；
    若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应 的音频段为语音信号。
  12. 根据权利要求11所述的终端设备，其特征在于，每帧所述音频子序列包含T个采样值；所述静音检测模型包括输入层、隐藏层及输出层，所述输入层包含T个输入节点，所述隐藏层由L层级联的降维网络构成，每层所述降维网络均配置有第一卷积核；
    所述将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值，包括：
    通过所述静音检测模型的输入层包含的T个输入节点分别接收所述音频子序列包含的T个采样值；
    在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组；
    在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值。
  13. 根据权利要求12所述的终端设备，其特征在于，所述第一卷积核的长度与其步长相等，每层所述降维网络接收到的音频子序列的长度为该层的第一卷积核的长度的整数倍；
    所述在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组，包括：
    在所述静音检测模型的隐藏层中，依次基于第一预设卷积公式对每层所述降维网络接收到的音频子序列进行卷积处理；所述第一预设卷积公式为：
    Figure PCTCN2019103149-appb-100011
    其中，Audio i为第i层所述降维网络待输出的音频子序列，i＝{1,2,…,L}，Kernel ij为第i层所述降维网络的第一卷积核中第j个元素的值，k i为Kernel ij的长度，Audio (i-1)j为第i-1层所述降维网络输出的音频子序列包含的第j个音频元素的值，
    Figure PCTCN2019103149-appb-100012
    为第i-1层所述降维网络输出的音频子序列包含的第j+k i个音频元素的值，
    Figure PCTCN2019103149-appb-100013
    为第i-1层 所述降维网络输出的音频子序列包含的第
    Figure PCTCN2019103149-appb-100014
    个音频元素的值，a i-1为第i-1层所述降维网络输出的音频子序列的长度；
    将第L层所述降维网络输出的卷积处理后的所述音频子序列确定为所述音频子序列的特征数组。
  14. 根据权利要求12所述的终端设备，其特征在于，所述第二卷积核的长度与所述音频子序列的特征数组的长度相等；
    所述在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值，包括：
    在所述输出层中基于第二预设卷积公式对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值；所述第二预设卷积公式为：
    Figure PCTCN2019103149-appb-100015
    其中，Audio final为所述音频子序列的特征值，a final为所述音频子序列的特征数组的长度，Kernel j为所述第二卷积核中第j个元素的值，Audio j为所述音频子序列的特征数组中第j个音频元素的值。
  15. 根据权利要求11-14任一项所述的终端设备，其特征在于，所述将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值之后，还包括：
    若所述音频子序列的特征值小于预设特征值阈值，则确定所述音频子序列对应的音频段为噪声信号。
  16. 一种计算机非易失性可读存储介质，所述计算机非易失性可读存储介质存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现如下步骤：
    基于预设采样频率对待检测的原始音频信号进行采样，得到所述原始音频信号对应的采样信号；
    基于预设感受野长度对所述采样信号进行分帧处理，得到至少两帧音频子序列；
    将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值；所述静音检测模型为一维卷积神经网络模型，所述音频子序列的特征值用于表征所述音频子序列对应的音频段为语音信号的概率，所述特征值为一维数值；
    若所述音频子序列的特征值大于或等于预设特征值阈值，则确定所述音频子序列对应的音频段为语音信号。
  17. 根据权利要求16所述的计算机非易失性可读存储介质，其特征在于，每帧所述音频子序列包含T个采样值；所述静音检测模型包括输入层、隐藏层及输出层，所述输入层包含T个输入节点，所述隐藏层由L层级联的降维网络构成，每层所述降维网络均配置有 第一卷积核；
    所述将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值，包括：
    通过所述静音检测模型的输入层包含的T个输入节点分别接收所述音频子序列包含的T个采样值；
    在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组；
    在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值。
  18. 根据权利要求17所述的计算机非易失性可读存储介质，其特征在于，所述第一卷积核的长度与其步长相等，每层所述降维网络接收到的音频子序列的长度为该层的第一卷积核的长度的整数倍；
    所述在所述静音检测模型的隐藏层中，依次基于每层所述降维网络的第一卷积核对每层所述降维网络接收到的音频子序列进行卷积处理，在第L层所述降维网络中得到所述音频子序列的特征数组，包括：
    在所述静音检测模型的隐藏层中，依次基于第一预设卷积公式对每层所述降维网络接收到的音频子序列进行卷积处理；所述第一预设卷积公式为：
    Figure PCTCN2019103149-appb-100016
    其中，Audio i为第i层所述降维网络待输出的音频子序列，i＝{1,2,…,L}，Kernel ij为第i层所述降维网络的第一卷积核中第j个元素的值，k i为Kernel ij的长度，Audio (i-1)j为第i-1层所述降维网络输出的音频子序列包含的第j个音频元素的值，
    Figure PCTCN2019103149-appb-100017
    为第i-1层所述降维网络输出的音频子序列包含的第j+k i个音频元素的值，
    Figure PCTCN2019103149-appb-100018
    为第i-1层所述降维网络输出的音频子序列包含的第
    Figure PCTCN2019103149-appb-100019
    个音频元素的值，a i-1为第i-1层所述 降维网络输出的音频子序列的长度；
    将第L层所述降维网络输出的卷积处理后的所述音频子序列确定为所述音频子序列的特征数组。
  19. 根据权利要求17所述的计算机非易失性可读存储介质，其特征在于，所述第二卷积核的长度与所述音频子序列的特征数组的长度相等；
    所述在所述静音检测模型的输出层基于第二卷积核对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值，包括：
    在所述输出层中基于第二预设卷积公式对所述音频子序列的特征数组进行卷积处理，得到所述音频子序列的特征值；所述第二预设卷积公式为：
    Figure PCTCN2019103149-appb-100020
    其中，Audio final为所述音频子序列的特征值，a final为所述音频子序列的特征数组的长度，Kernel j为所述第二卷积核中第j个元素的值，Audio j为所述音频子序列的特征数组中第j个音频元素的值。
  20. 根据权利要求16-19任一项所述的计算机非易失性可读存储介质，其特征在于，所述将所述音频子序列输入至预先训练好的静音检测模型，得到所述音频子序列的特征值之后，还包括：
    若所述音频子序列的特征值小于预设特征值阈值，则确定所述音频子序列对应的音频段为噪声信号。
PCT/CN2019/103149 2019-03-25 2019-08-29 一种基于神经网络的静音检测方法、终端设备及介质 Ceased WO2020192009A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910226470.2 2019-03-25
CN201910226470.2A CN110010153A (zh) 2019-03-25 2019-03-25 一种基于神经网络的静音检测方法、终端设备及介质

Publications (1)

Publication Number Publication Date
WO2020192009A1 true WO2020192009A1 (zh) 2020-10-01

Family

ID=67167950

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/103149 Ceased WO2020192009A1 (zh) 2019-03-25 2019-08-29 一种基于神经网络的静音检测方法、终端设备及介质

Country Status (2)

Country Link
CN (1) CN110010153A (zh)
WO (1) WO2020192009A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116417015A (zh) * 2023-04-03 2023-07-11 广州市迪士普音响科技有限公司 一种压缩音频的静默检测方法及装置

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110010153A (zh) * 2019-03-25 2019-07-12 平安科技（深圳）有限公司 一种基于神经网络的静音检测方法、终端设备及介质
CN111181949B (zh) * 2019-12-25 2023-12-12 视联动力信息技术股份有限公司 一种声音检测方法、装置、终端设备和存储介质
CN114446291B (zh) * 2020-11-04 2025-09-19 阿里巴巴集团控股有限公司 语音识别方法、装置、智能音箱、家电、电子设备及介质
CN114694636B (zh) 2020-12-31 2025-12-16 华为技术有限公司 语音识别方法及装置
CN113889086B (zh) * 2021-09-15 2025-03-25 青岛信芯微电子科技股份有限公司 语音识别模型的训练方法、语音识别方法及相关装置
CN116469413B (zh) * 2023-04-03 2023-12-01 广州市迪士普音响科技有限公司 一种基于人工智能的压缩音频静默检测方法及装置

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2001086633A1 (en) * 2000-05-10 2001-11-15 Multimedia Technologies Institute - Mti S.R.L. Voice activity detection and end-point detection
CN102693724A (zh) * 2011-03-22 2012-09-26 张燕 一种基于神经网络的高斯混合模型的噪声分类方法
CN105427870A (zh) * 2015-12-23 2016-03-23 北京奇虎科技有限公司 一种针对停顿的语音识别方法和装置
CN107393526A (zh) * 2017-07-19 2017-11-24 腾讯科技（深圳）有限公司 语音静音检测方法、装置、计算机设备和存储介质
US20180166066A1 (en) * 2016-12-14 2018-06-14 International Business Machines Corporation Using long short-term memory recurrent neural network for speaker diarization segmentation
CN108428448A (zh) * 2017-02-13 2018-08-21 芋头科技（杭州）有限公司 一种语音端点检测方法及语音识别方法
CN109036467A (zh) * 2018-10-26 2018-12-18 南京邮电大学 基于tf-lstm的cffd提取方法、语音情感识别方法及系统
CN109146066A (zh) * 2018-11-01 2019-01-04 重庆邮电大学 一种基于语音情感识别的虚拟学习环境自然交互方法
CN110010153A (zh) * 2019-03-25 2019-07-12 平安科技（深圳）有限公司 一种基于神经网络的静音检测方法、终端设备及介质

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10229700B2 (en) * 2015-09-24 2019-03-12 Google Llc Voice activity detection
US11080591B2 (en) * 2016-09-06 2021-08-03 Deepmind Technologies Limited Processing sequences using convolutional neural networks
CN108346433A (zh) * 2017-12-28 2018-07-31 北京搜狗科技发展有限公司 一种音频处理方法、装置、设备及可读存储介质
CN109036459B (zh) * 2018-08-22 2019-12-27 百度在线网络技术（北京）有限公司 语音端点检测方法、装置、计算机设备、计算机存储介质
CN109378016A (zh) * 2018-10-10 2019-02-22 四川长虹电器股份有限公司 一种基于vad的关键词识别标注方法

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2001086633A1 (en) * 2000-05-10 2001-11-15 Multimedia Technologies Institute - Mti S.R.L. Voice activity detection and end-point detection
CN102693724A (zh) * 2011-03-22 2012-09-26 张燕 一种基于神经网络的高斯混合模型的噪声分类方法
CN105427870A (zh) * 2015-12-23 2016-03-23 北京奇虎科技有限公司 一种针对停顿的语音识别方法和装置
US20180166066A1 (en) * 2016-12-14 2018-06-14 International Business Machines Corporation Using long short-term memory recurrent neural network for speaker diarization segmentation
CN108428448A (zh) * 2017-02-13 2018-08-21 芋头科技（杭州）有限公司 一种语音端点检测方法及语音识别方法
CN107393526A (zh) * 2017-07-19 2017-11-24 腾讯科技（深圳）有限公司 语音静音检测方法、装置、计算机设备和存储介质
CN109036467A (zh) * 2018-10-26 2018-12-18 南京邮电大学 基于tf-lstm的cffd提取方法、语音情感识别方法及系统
CN109146066A (zh) * 2018-11-01 2019-01-04 重庆邮电大学 一种基于语音情感识别的虚拟学习环境自然交互方法
CN110010153A (zh) * 2019-03-25 2019-07-12 平安科技（深圳）有限公司 一种基于神经网络的静音检测方法、终端设备及介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116417015A (zh) * 2023-04-03 2023-07-11 广州市迪士普音响科技有限公司 一种压缩音频的静默检测方法及装置
CN116417015B (zh) * 2023-04-03 2023-09-12 广州市迪士普音响科技有限公司 一种压缩音频的静默检测方法及装置

Also Published As

Publication number Publication date
CN110010153A (zh) 2019-07-12

Similar Documents

Publication Publication Date Title
WO2020192009A1 (zh) 一种基于神经网络的静音检测方法、终端设备及介质
WO2022141868A1 (zh) 一种提取语音特征的方法、装置、终端及存储介质
CN108962237B (zh) 混合语音识别方法、装置及计算机可读存储介质
CN112183107B (zh) 音频的处理方法和装置
WO2022121257A1 (zh) 模型训练方法、语音识别方法、装置、设备及存储介质
WO2022121180A1 (zh) 模型的训练方法、装置、语音转换方法、设备及存储介质
CN110364185B (zh) 一种基于语音数据的情绪识别方法、终端设备及介质
WO2021189642A1 (zh) 信号处理方法、装置、计算机设备及存储介质
CN108922543B (zh) 模型库建立方法、语音识别方法、装置、设备及介质
WO2020237869A1 (zh) 一种问题意图识别方法、装置、计算机设备及存储介质
CN106952649A (zh) 基于卷积神经网络和频谱图的说话人识别方法
CN113646833A (zh) 语音对抗样本检测方法、装置、设备及计算机可读存储介质
WO2022178942A1 (zh) 情绪识别方法、装置、计算机设备和存储介质
WO2020098256A1 (zh) 基于全卷积神经网络的语音增强方法、装置及存储介质
WO2020151338A1 (zh) 一种音频噪声的检测方法、装置、存储介质和移动终端
CN113160795B (zh) 语种特征提取模型训练方法、装置、设备及存储介质
WO2020140609A1 (zh) 一种语音识别方法、设备及计算机可读存储介质
WO2022121188A1 (zh) 关键词检测方法、装置、设备和存储介质
CN115329705A (zh) 半导体器件的布局方法、装置、可读存储介质、终端
CN112885367B (zh) 基频获取方法、装置、计算机设备和存储介质
WO2021051507A1 (zh) 一种机器人对话生成方法、装置、可读存储介质及机器人
CN112712792B (zh) 一种方言识别模型的训练方法、可读存储介质及终端设备
WO2025061091A1 (zh) 语音唤醒模型的训练和执行方法、装置、设备及存储介质
CN118098206B (zh) 命令词得分计算方法、装置、设备和介质
CN114743561B (zh) 语音分离装置及方法、存储介质、计算机设备

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: 19921058

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19921058

Country of ref document: EP

Kind code of ref document: A1