WO2020238046A1 - Human voice smart detection method and apparatus, and computer readable storage medium - Google Patents

Human voice smart detection method and apparatus, and computer readable storage medium Download PDF

Info

Publication number
WO2020238046A1
WO2020238046A1 PCT/CN2019/117352 CN2019117352W WO2020238046A1 WO 2020238046 A1 WO2020238046 A1 WO 2020238046A1 CN 2019117352 W CN2019117352 W CN 2019117352W WO 2020238046 A1 WO2020238046 A1 WO 2020238046A1
Authority
WO
WIPO (PCT)
Prior art keywords
training
human voice
data
input
emphasis
Prior art date
Application number
PCT/CN2019/117352
Other languages
French (fr)
Chinese (zh)
Inventor
王健宗
程宁
Original Assignee
平安科技（深圳）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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2020238046A1 publication Critical patent/WO2020238046A1/en

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
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/02—Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
    • 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
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/04—Training, enrolment or model building
    • 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
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/18—Artificial neural networks; Connectionist approaches
    • 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
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/26—Recognition of special voice characteristics, e.g. for use in lie detectors; Recognition of animal voices

Definitions

  • This application relates to the field of artificial intelligence technology, and in particular to a method, device and computer-readable storage medium that can intelligently detect whether there is a human voice based on voice data input.
  • Video surveillance systems have been widely used, but most video surveillance systems currently do not detect human voices. Furthermore, the main content of research in the field of human voice detection at home and abroad includes the recognition of voice features of different people, voice recognition of different semantic features, and voice recognition of different emotional state features, but the commonality of most research is that people are known to be human. Under the premise of the spoken voice, to study a certain aspect of the human voice, there are few studies that directly detect whether it is a human voice, and due to the variability between human voice and the environment, most human voices The detection method is not ideal in practical applications, and the effect of human voice detection needs to be resolved in time.
  • This application provides a human voice intelligent detection method, device, and computer-readable storage medium, the main purpose of which is to determine whether the voice data includes accurate results of human voice when the user inputs voice data.
  • a human voice intelligent detection method includes:
  • the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set includes pre-processing. Preprocessing operations of emphasis and windowing and framing, input the training set completed by the preprocessing operation to the human voice detection model, and input the label set to the loss function;
  • the human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the training value
  • the loss value is calculated, and the size of the loss value and the preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
  • the input voice data is received and input to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
  • performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
  • H(z) is the training set after the pre-emphasis
  • z is the sound frequency
  • ⁇ is the pre-emphasis coefficient
  • n is the training set after the pre-emphasis
  • N is the window length of the Hamming window method
  • cos is the cosine function
  • performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
  • H(z) is the training set after the pre-emphasis
  • z is the sound frequency
  • ⁇ is the pre-emphasis coefficient
  • n is the training set after the pre-emphasis
  • N is the window length of the Hamming window method
  • cos is the cosine function
  • the human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, including:
  • the first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
  • the fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
  • the convolution operation is:
  • ⁇ ' is output data
  • ⁇ is input data
  • k is the size of the convolution kernel
  • s is the step size of the convolution operation
  • p is the data zero-filling matrix
  • the activation function is:
  • the present application also provides a human voice intelligent detection device, which includes a memory and a processor.
  • the memory stores a human voice intelligent detection program that can run on the processor.
  • the human voice intelligent detection program is executed by the processor, the following steps are implemented:
  • the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set includes pre-processing. Preprocessing operations of emphasis and windowing and framing, input the training set completed by the preprocessing operation to the human voice detection model, and input the label set to the loss function;
  • the human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the training value
  • the loss value is calculated, and the size of the loss value and the preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
  • the input voice data is received and input to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
  • performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
  • H(z) is the training set after the pre-emphasis
  • z is the sound frequency
  • ⁇ is the pre-emphasis coefficient
  • n is the training set after the pre-emphasis
  • N is the window length of the Hamming window method
  • cos is the cosine function
  • performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
  • H(z) is the training set after the pre-emphasis
  • z is the sound frequency
  • ⁇ is the pre-emphasis coefficient
  • n is the training set after the pre-emphasis
  • N is the window length of the Hamming window method
  • cos is the cosine function
  • the human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, including:
  • the first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
  • the fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
  • the present application also provides a computer-readable storage medium, the computer-readable storage medium stores a human voice intelligent detection program, and the human voice intelligent detection program can be executed by one or more processors , In order to realize the steps of the human voice intelligent detection method described above.
  • the human voice detection model of the present application uses a convolutional neural network.
  • the convolutional neural network retains the associated information between voices based on the idea of local perception and weight sharing, which can greatly reduce the number of required parameters and is further improved by pooling. The number of network parameters is reduced and the robustness of the model is improved. Therefore, the human voice intelligent detection method, device, and computer-readable storage medium proposed in this application can realize efficient human voice detection judgment.
  • FIG. 1 is a schematic flowchart of a human voice intelligent detection method provided by an embodiment of the application
  • FIG. 2 is a schematic diagram of the internal structure of a human voice intelligent detection device provided by an embodiment of the application;
  • FIG. 3 is a schematic diagram of modules of a human voice intelligent detection program in a human voice intelligent detection device provided by an embodiment of the application.
  • This application provides a human voice intelligent detection method.
  • FIG. 1 it is a schematic flowchart of a human voice intelligent detection method provided by an embodiment of this application.
  • the method can be executed by a device, and the device can be implemented by software and/or hardware.
  • the human voice intelligent detection method includes:
  • the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set is performed
  • the pre-processing operations include pre-emphasis and windowing and framing.
  • the training set completed by the pre-processing operation is input to the human voice detection model, and the label set is input to the loss function.
  • the positive sample set including human voice data is recorded by a microphone in a quiet environment
  • the sampling frequency of the microphone recording is 16kHz
  • the sampling size is 16bits
  • the persons participating in the admission record at least two different sections
  • the vocal data, one section is admitted in standard Mandarin, and the other section is admitted in the local dialect of the admitted person.
  • the duration of each piece of human voice data in the positive sample set is not less than 10 seconds.
  • the negative sample set comes from the audio data set AudioSet, which includes multiple manually marked sound clips.
  • the AudioSet is a large-scale and complete audio data set that is currently open.
  • the Multiple manually marked sound clips include 2084320 manually marked sound clips each with a length of 10 seconds.
  • the preferred implementation of the pre-emphasis pre-processing operation in this application is to improve the high-frequency range part of the training set, so that the signal spectrum from the low-frequency range to the high-frequency range of the training set becomes flat, and at the same time, it can suppress random noise and Further, the pre-emphasis is based on the digital filter to pre-emphasize the sound frequency of the training set, and the method of the pre-emphasis, that is, the pre-emphasis, is:
  • H(z) is the training set after the pre-emphasis
  • z is the sound frequency
  • ⁇ is the pre-emphasis coefficient
  • the preferred implementation of the windowing and framing in this application is based on the feature that the audio signal of the training set remains unchanged within a small range of time, and the audio signal of the training set is subjected to framing processing, and further, The windowing and framing is based on the pre-emphasized training set, and the windowing and framing processing is performed according to the Hamming window method, and the Hamming window method ⁇ (n) is:
  • n is the training set after the pre-emphasis
  • N is the window length of the Hamming window method
  • cos is the cosine function
  • the human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the The training value is calculated to obtain a loss value, and the size of the loss value and a preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits the training.
  • the human voice detection model in the preferred embodiment of the present application receives the training set completed by the preprocessing operation, and inputs the training set to the first layer of convolutional layer.
  • the first layer of convolutional layer is subjected to convolution, Obtain a convolutional data set and input it to the first layer of pooling layer; then, the first layer of pooling layer performs a maximization pooling operation on the convolutional data set to obtain a dimensionality reduction data set and input it to the second layer of volume Multilayer, the second layer of convolutional layer performs the convolution operation and then inputs to the second layer of pooling layer for the maximum pooling operation, until finally input to the fully connected layer; the fully connected layer is combined with activation Function calculation to obtain the training value;
  • ⁇ is the output data
  • ⁇ is the input data
  • k is the size of the convolution kernel
  • s is the step size of the convolution operation
  • p is the data zero-filling matrix
  • n is the size of the training set
  • y t is the training value
  • ⁇ t is the label set.
  • the human voice detection model judges whether the sound data includes human voice and outputs a judgment result.
  • the invention also provides a human voice intelligent detection device.
  • FIG. 2 it is a schematic diagram of the internal structure of a human voice intelligent detection device provided by an embodiment of this application.
  • the human voice intelligent detection device 1 may be a PC (Personal Computer, personal computer), or a terminal device such as a smart phone, a tablet computer, or a portable computer, or a server.
  • the human voice intelligent detection device 1 at least includes a memory 11, a processor 12, a communication bus 13, and a network interface 14.
  • the memory 11 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, etc.
  • the memory 11 may be an internal storage unit of the human voice intelligent detection device 1 in some embodiments, such as a hard disk of the human voice intelligent detection device 1.
  • the memory 11 may also be an external storage device of the human voice intelligent detection device 1, for example, a plug-in hard disk equipped on the human voice intelligent detection device 1, a smart media card (SMC), and a secure digital (Secure Digital, SD) card, flash card (Flash Card), etc.
  • the memory 11 may also include both an internal storage unit of the human voice intelligent detection device 1 and an external storage device.
  • the memory 11 can be used not only to store application software and various data installed in the human voice intelligent detection device 1, such as the code of the human voice intelligent detection program 01, etc., but also to temporarily store data that has been output or will be output.
  • the processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip, and is used to run the program code or processing stored in the memory 11 Data, for example, execute the human voice intelligent detection program 01 (the human voice intelligent detection program 01 is essentially a software system) and so on.
  • CPU central processing unit
  • controller microcontroller
  • microprocessor or other data processing chip
  • the communication bus 13 is used to realize the connection and communication between these components.
  • the network interface 14 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 device 1 and other electronic devices.
  • the device 1 may also include a user interface.
  • the user interface may include a display (Display) and an input unit such as a keyboard (Keyboard).
  • the optional user interface may also include a standard wired interface and a wireless interface.
  • the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light emitting diode) touch device, etc.
  • the display may also be appropriately called a display screen or a display unit, which is used to display the information processed in the human voice intelligent detection device 1 and to display a visualized user interface.
  • Figure 2 only shows the human voice intelligent detection device 1 with components 11-14 and human voice intelligent detection program 01. Those skilled in the art will understand that the structure shown in Figure 1 does not constitute a human voice intelligent detection device
  • the definition of 1 may include fewer or more components than shown, or a combination of certain components, or different component arrangements.
  • the human voice intelligent detection program 01 is stored in the memory 11; when the processor 12 executes the human voice intelligent detection program 01 stored in the memory 11, the following steps are implemented:
  • Step 1 The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set A pre-processing operation including pre-emphasis and windowing and framing is performed, the training set completed by the pre-processing operation is input to the human voice detection model, and the label set is input to the loss function.
  • the positive sample set including human voice data is recorded by a microphone in a quiet environment
  • the sampling frequency of the microphone recording is 16kHz
  • the sampling size is 16bits
  • the persons participating in the admission record at least two different sections
  • the vocal data, one section is admitted in standard Mandarin, and the other section is admitted in the local dialect of the admitted person.
  • the duration of each segment of human voice data in the positive sample set is not less than 10 seconds.
  • the negative sample set comes from the audio data set AudioSet, which includes multiple manually marked sound clips.
  • the AudioSet is a large-scale and complete audio data set that is currently open.
  • the Multiple manually marked sound clips include 2084320 manually marked sound clips each with a length of 10 seconds.
  • the preferred implementation of the pre-emphasis pre-processing operation in this application is to improve the high-frequency range part of the training set, so that the signal spectrum from the low-frequency range to the high-frequency range of the training set becomes flat, and at the same time, it can suppress random noise and Further, the pre-emphasis is based on the digital filter to pre-emphasize the sound frequency of the training set, and the method of the pre-emphasis, that is, the pre-emphasis, is:
  • H(z) is the training set after the pre-emphasis
  • z is the sound frequency
  • ⁇ is the pre-emphasis coefficient
  • the preferred implementation of the windowing and framing in this application is based on the feature that the audio signal of the training set remains unchanged within a small range of time, and the audio signal of the training set is subjected to framing processing, and further, The windowing and framing is based on the pre-emphasized training set, and the windowing and framing processing is performed according to the Hamming window method, and the Hamming window method ⁇ (n) is:
  • n is the training set after the pre-emphasis
  • N is the window length of the Hamming window method
  • cos is the cosine function
  • Step 2 The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the The training value is calculated to obtain a loss value, and the size of the loss value and a preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits the training.
  • the human voice detection model described in the preferred embodiment of the present application receives the training set completed by the preprocessing operation, and inputs the training set to the first layer of convolutional layer.
  • the first layer of convolutional layer performs convolution operation Obtain the convolutional data set and input it to the first-layer pooling layer; then, the first-layer pooling layer performs a maximization pooling operation on the convolutional data set to obtain a dimensionality reduction data set and input it to the second-layer volume Layers, the second convolutional layer performs the convolution operation and then inputs to the second pooling layer for the maximization pooling operation, until finally input to the fully connected layer; the fully connected layer is combined with activation Function calculation to obtain the training value;
  • ⁇ is the output data
  • ⁇ is the input data
  • k is the size of the convolution kernel
  • s is the step size of the convolution operation
  • p is the data zero-filling matrix
  • n is the size of the training set
  • y t is the training value
  • ⁇ t is the label set.
  • Step 3 The input voice data is received and input to the human voice detection model, and the human voice detection model judges whether the voice data includes human voice and outputs the judgment result.
  • the human voice intelligent detection program can also be divided into one or more modules, and the one or more modules are stored in the memory 11 and run by one or more processors (this embodiment It is 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, and is used to describe the execution process of the human voice intelligent detection program in the human voice intelligent detection device .
  • FIG. 3 it is a schematic diagram of the program modules of the human voice intelligent detection program in an embodiment of the applicant's voice intelligent detection device.
  • the human voice intelligent detection program can be divided into data receiving modules 10.
  • the data receiving module 10 is configured to receive a positive sample set including human voice data, a negative sample set not including human voice data, and a label set.
  • the positive sample set and the negative sample set are collectively referred to as a training set, and the The training set is subjected to pre-processing operations including pre-emphasis and windowing and framing, the training set completed by the pre-processing operation is input to the human voice detection model, and the label set is input to the loss function.
  • the model training module 20 is configured to: the human voice detection model receives the training set completed by the preprocessing operation to obtain training values, and inputs the training values into the loss function, which is based on the The label set and the training value are calculated to obtain a loss value, and the size of the loss value and a preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training.
  • the human voice result output module 30 is configured to receive input voice data and input it to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
  • the embodiment of the present application also proposes a computer-readable storage medium.
  • the computer-readable storage medium stores a human voice intelligent detection program, and the human voice intelligent detection program can be executed by one or more processors to realize Do as follows:
  • the pre-processing operation of framing is to input the training set completed by the pre-processing operation into the human voice detection model, and input the label set into the loss function.
  • the human voice detection model receives the training set completed by the preprocessing operation for training to obtain training values, and inputs the training values to the loss function, and the loss function is calculated based on the label set and the training value The loss value is obtained, and the size of the loss value and a preset threshold is judged, and the human voice detection model exits training when the loss value is less than the preset threshold.
  • the input voice data is received and input to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Image Analysis (AREA)
  • Electrically Operated Instructional Devices (AREA)

Abstract

The present application relates to artificial intelligence technology, and discloses a human voice smart detection method, comprising: receiving a label set and a training set comprising a positive sample set and a negative sample set and, after performing pre-processing operations including pre-emphasis and windowing and framing on the training set, inputting same into a human voice detection model, and inputting the label set into a loss function; the human voice detection model receives the pre-processed training set, performs training to obtain a training value, inputs the training value into the loss function, the loss function being calculated to obtain a loss value, and determines the size of the loss value and a preset threshold value until the loss value is less than the preset threshold value, and then the human voice detection model exits training; receiving inputted audio data, using the human voice detection model to determine whether the audio data comprises a human voice, and outputting a determining result. Also provided in the present application are a human voice smart detection apparatus and a computer readable storage medium. The present application can implement highly effective human voice detection.

Description

人声智能检测方法、装置及计算机可读存储介质Human voice intelligent detection method, device and computer readable storage medium
本申请要求于2019年05月29日提交中国专利局、申请号为201910468133.4、发明名称为“人声智能检测方法、装置及计算机可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在申请中。This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on May 29, 2019, the application number is 201910468133.4, and the invention title is "Human Voice Intelligent Detection Method, Device and Computer-readable Storage Medium". The reference is incorporated in the application.
技术领域Technical field
本申请涉及人工智能技术领域，尤其涉及一种基于语音数据输入后可智能化检测是否有人声的方法、装置及计算机可读存储介质。This application relates to the field of artificial intelligence technology, and in particular to a method, device and computer-readable storage medium that can intelligently detect whether there is a human voice based on voice data input.
背景技术Background technique
视频监控系统目前已得到广泛的应用，然而目前多数视频监控系统没有对人声进行检测。进一步地，国内外的人声检测领域研究的主要内容包括识别不同人的声音特征以及不同语义特征的人声识别和不同情感状态特征的人声识别等，但多数研究的共性是已知是人所发出的说话声音的前提下，研究所述人声的某一方面特征，很少有直接对是否是人声进行检测的研究，且由于人声与环境之间的多变性，使得多数人声检测方法在实际应用中效果不理想，人声检测的效果有待及时解决。Video surveillance systems have been widely used, but most video surveillance systems currently do not detect human voices. Furthermore, the main content of research in the field of human voice detection at home and abroad includes the recognition of voice features of different people, voice recognition of different semantic features, and voice recognition of different emotional state features, but the commonality of most research is that people are known to be human. Under the premise of the spoken voice, to study a certain aspect of the human voice, there are few studies that directly detect whether it is a human voice, and due to the variability between human voice and the environment, most human voices The detection method is not ideal in practical applications, and the effect of human voice detection needs to be resolved in time.
发明内容Summary of the invention
本申请提供一种人声智能检测方法、装置及计算机可读存储介质，其主要目的在于当用户输入语音数据时，给用户判断所述语音数据是否包括人声的精准结果。This application provides a human voice intelligent detection method, device, and computer-readable storage medium, the main purpose of which is to determine whether the voice data includes accurate results of human voice when the user inputs voice data.
为实现上述目的，本申请提供的一种人声智能检测方法，包括：In order to achieve the above objective, a human voice intelligent detection method provided by this application includes:
数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数；The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set includes pre-processing. Preprocessing operations of emphasis and windowing and framing, input the training set completed by the preprocessing operation to the human voice detection model, and input the label set to the loss function;
所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集 和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练；The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the training value The loss value is calculated, and the size of the loss value and the preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
接收输入的声音数据并输入至所述人声检测模型，所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The input voice data is received and input to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
可选地，对所述训练集进行包括预加重和加窗分帧的预处理操作，包括：Optionally, performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
H(z)＝1-μz -1 H(z) = 1-μz -1
其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
Figure PCTCN2019117352-appb-000001
Figure PCTCN2019117352-appb-000001
其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
可选地，对所述训练集进行包括预加重和加窗分帧的预处理操作，包括：Optionally, performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
H(z)＝1-μz -1 H(z) = 1-μz -1
其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
Figure PCTCN2019117352-appb-000002
Figure PCTCN2019117352-appb-000002
其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
可选地，Optionally,
所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，包括：The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, including:
将所述训练集输入至所述人声检测模型的第一层卷积层进行卷积操作，得到第一卷积数据集，并将所述第一卷积数据集输入至第一层池化层；Input the training set to the first convolutional layer of the human voice detection model to perform a convolution operation to obtain a first convolutional data set, and input the first convolutional data set to the first layer of pooling Floor;
所述第一层池化层对所述第一卷积数据集进行最大化池化操作，得到第 一降维数据集，并将所述第一降维数据集输入至第二层卷积层进行所述卷积操作，得到第二卷积数据集，将所述第二卷积数据集输入至第二层池化层进行所述最大化池化操作，得到第二降维数据集，并将所述第二降维数据集输入至全连接层；The first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
所述全连接层结合激活函数对所述第二降维数据集执行计算，得到所述训练值。The fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
可选地，所述卷积操作为：Optionally, the convolution operation is:
Figure PCTCN2019117352-appb-000003
Figure PCTCN2019117352-appb-000003
其中ω’为输出数据，ω为输入数据，k为卷积核的大小，s为所述卷积操作的步幅，p为数据补零矩阵；Where ω'is output data, ω is input data, k is the size of the convolution kernel, s is the step size of the convolution operation, and p is the data zero-filling matrix;
所述激活函数为：The activation function is:
Figure PCTCN2019117352-appb-000004
Figure PCTCN2019117352-appb-000004
其中y为所述第二降维数据集，e为无限不循环小数。Where y is the second dimensionality reduction data set, and e is an infinite non-cyclic decimal.
此外，为实现上述目的，本申请还提供一种人声智能检测装置，该装置包括存储器和处理器，所述存储器中存储有可在所述处理器上运行的人声智能检测程序，所述人声智能检测程序被所述处理器执行时实现如下步骤：In addition, in order to achieve the above-mentioned object, the present application also provides a human voice intelligent detection device, which includes a memory and a processor. The memory stores a human voice intelligent detection program that can run on the processor. When the human voice intelligent detection program is executed by the processor, the following steps are implemented:
数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数；The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set includes pre-processing. Preprocessing operations of emphasis and windowing and framing, input the training set completed by the preprocessing operation to the human voice detection model, and input the label set to the loss function;
所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练；The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the training value The loss value is calculated, and the size of the loss value and the preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
接收输入的声音数据并输入至所述人声检测模型，所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The input voice data is received and input to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
可选地，对所述训练集进行包括预加重和加窗分帧的预处理操作，包括：Optionally, performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
H(z)＝1-μz -1 H(z) = 1-μz -1
其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
Figure PCTCN2019117352-appb-000005
Figure PCTCN2019117352-appb-000005
其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
可选地，对所述训练集进行包括预加重和加窗分帧的预处理操作，包括：Optionally, performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
H(z)＝1-μz -1 H(z) = 1-μz -1
其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
Figure PCTCN2019117352-appb-000006
Figure PCTCN2019117352-appb-000006
其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
可选地，Optionally,
所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，包括：The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, including:
将所述训练集输入至所述人声检测模型的第一层卷积层进行卷积操作，得到第一卷积数据集，并将所述第一卷积数据集输入至第一层池化层；Input the training set to the first convolutional layer of the human voice detection model to perform a convolution operation to obtain a first convolutional data set, and input the first convolutional data set to the first layer of pooling Floor;
所述第一层池化层对所述第一卷积数据集进行最大化池化操作，得到第一降维数据集，并将所述第一降维数据集输入至第二层卷积层进行所述卷积操作，得到第二卷积数据集，将所述第二卷积数据集输入至第二层池化层进行所述最大化池化操作，得到第二降维数据集，并将所述第二降维数据集输入至全连接层；The first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
所述全连接层结合激活函数对所述第二降维数据集执行计算，得到所述训练值。The fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
此外，为实现上述目的，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上存储有人声智能检测程序，所述人声智能检测程序可被一个或者多个处理器执行，以实现如上所述的人声智能检测方法的步骤。In addition, in order to achieve the above object, the present application also provides a computer-readable storage medium, the computer-readable storage medium stores a human voice intelligent detection program, and the human voice intelligent detection program can be executed by one or more processors , In order to realize the steps of the human voice intelligent detection method described above.
本申请的人声检测模型使用卷积神经网络，所述卷积神经网络基于局部感知和权值共享思想保留了语音间的关联信息，可大大减少所需参数的数量，且通过池化操作进一步缩减网络参数数量，提高模型的鲁棒性，因此本申请提出的人声智能检测方法、装置及计算机可读存储介质可以实现高效的人声检测判断。The human voice detection model of the present application uses a convolutional neural network. The convolutional neural network retains the associated information between voices based on the idea of local perception and weight sharing, which can greatly reduce the number of required parameters and is further improved by pooling. The number of network parameters is reduced and the robustness of the model is improved. Therefore, the human voice intelligent detection method, device, and computer-readable storage medium proposed in this application can realize efficient human voice detection judgment.
附图说明Description of the drawings
图1为本申请一实施例提供的人声智能检测方法的流程示意图；FIG. 1 is a schematic flowchart of a human voice intelligent detection method provided by an embodiment of the application;
图2为本申请一实施例提供的人声智能检测装置的内部结构示意图；2 is a schematic diagram of the internal structure of a human voice intelligent detection device provided by an embodiment of the application;
图3为本申请一实施例提供的人声智能检测装置中人声智能检测程序的模块示意图。3 is a schematic diagram of modules of a human voice intelligent detection program in a human voice intelligent detection device provided by an embodiment of the application.
本申请目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。The realization, functional characteristics, and advantages of the purpose of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings.
具体实施方式Detailed ways
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。It should be understood that the specific embodiments described here are only used to explain the application, and are not used to limit the application.
本申请提供一种人声智能检测方法。参照图1所示，为本申请一实施例提供的人声智能检测方法的流程示意图。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。This application provides a human voice intelligent detection method. Referring to FIG. 1, it is a schematic flowchart of a human voice intelligent detection method provided by an embodiment of this application. The method can be executed by a device, and the device can be implemented by software and/or hardware.
在本实施例中，人声智能检测方法包括：In this embodiment, the human voice intelligent detection method includes:
S1、数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数。S1. The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set is performed The pre-processing operations include pre-emphasis and windowing and framing. The training set completed by the pre-processing operation is input to the human voice detection model, and the label set is input to the loss function.
本申请较佳实施例，所述包括人声数据的正样本集是在安静的环境下通过麦克风录取，所述麦克风录取的采样频率16kHz、采样大小16bits，且参与录取的人员至少录制两段不同人声数据，一段使用标准普通话录取，另一段 使用所述录取人员的地方方言录取。所述正样本集内每段人声数据时长不少于10秒。In a preferred embodiment of the present application, the positive sample set including human voice data is recorded by a microphone in a quiet environment, the sampling frequency of the microphone recording is 16kHz, the sampling size is 16bits, and the persons participating in the admission record at least two different sections The vocal data, one section is admitted in standard Mandarin, and the other section is admitted in the local dialect of the admitted person. The duration of each piece of human voice data in the positive sample set is not less than 10 seconds.
本申请较佳实施例，所述负样本集来源于音频数据集AudioSet中包括多条人工标记的声音剪辑片段，所述AudioSet是目前开放的大规模且完善的音频数据集，进一步地，所述多条人工标记的声音剪辑片段包括2084320条人工标记的每段10秒长度的声音剪辑片段。In a preferred embodiment of the present application, the negative sample set comes from the audio data set AudioSet, which includes multiple manually marked sound clips. The AudioSet is a large-scale and complete audio data set that is currently open. Further, the Multiple manually marked sound clips include 2084320 manually marked sound clips each with a length of 10 seconds.
本申请较佳实施所述预加重预处理操作是提高所述训练集的高频音域部分，使所述训练集的音域低频到音域高频的信号频谱变得平坦，同时还能抑制随机噪声和直流漂移的影响，进一步地，所述预加重是基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重即所述预加重的方法为：The preferred implementation of the pre-emphasis pre-processing operation in this application is to improve the high-frequency range part of the training set, so that the signal spectrum from the low-frequency range to the high-frequency range of the training set becomes flat, and at the same time, it can suppress random noise and Further, the pre-emphasis is based on the digital filter to pre-emphasize the sound frequency of the training set, and the method of the pre-emphasis, that is, the pre-emphasis, is:
H(z)＝1-μz -1 H(z) = 1-μz -1
其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
本申请较佳实施所述加窗分帧是根据在小范围的时间内，所述训练集的音频信号保持不变的特点，对所述训练集的音频信号进行分帧处理，进一地，所述加窗分帧基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：The preferred implementation of the windowing and framing in this application is based on the feature that the audio signal of the training set remains unchanged within a small range of time, and the audio signal of the training set is subjected to framing processing, and further, The windowing and framing is based on the pre-emphasized training set, and the windowing and framing processing is performed according to the Hamming window method, and the Hamming window method ω(n) is:
Figure PCTCN2019117352-appb-000007
Figure PCTCN2019117352-appb-000007
其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
S2、所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练。S2. The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the The training value is calculated to obtain a loss value, and the size of the loss value and a preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits the training.
本申请较佳实施例所述人声检测模型接收所述预处理操作完成的训练集，将所述训练集输入至第一层卷积层，所述第一层卷积层进行卷积操作后得到卷积数据集输入至第一层池化层；其后，所述第一层池化层对所述卷积数据集进行最大化池化操作后得到降维数据集输入至第二层卷积层，所述第二层卷积层进行所述卷积操作后输入至第二层池化层进行所述最大化池化操作，直到最终输入至全连接层；所述全连接层结合激活函数计算得到所述训练值；The human voice detection model in the preferred embodiment of the present application receives the training set completed by the preprocessing operation, and inputs the training set to the first layer of convolutional layer. After the first layer of convolutional layer is subjected to convolution, Obtain a convolutional data set and input it to the first layer of pooling layer; then, the first layer of pooling layer performs a maximization pooling operation on the convolutional data set to obtain a dimensionality reduction data set and input it to the second layer of volume Multilayer, the second layer of convolutional layer performs the convolution operation and then inputs to the second layer of pooling layer for the maximum pooling operation, until finally input to the fully connected layer; the fully connected layer is combined with activation Function calculation to obtain the training value;
本申请较佳实施例所述卷积操作为：The convolution operation described in the preferred embodiment of this application is:
Figure PCTCN2019117352-appb-000008
Figure PCTCN2019117352-appb-000008
其中ω’为输出数据，ω为输入数据，k为卷积核的大小，s为卷积操作的步幅，p为数据补零矩阵；Where ω’ is the output data, ω is the input data, k is the size of the convolution kernel, s is the step size of the convolution operation, and p is the data zero-filling matrix;
本申请较佳实施例所述激活函数为：The activation function described in the preferred embodiment of this application is:
Figure PCTCN2019117352-appb-000009
Figure PCTCN2019117352-appb-000009
其中y为所述第二降维数据集，e为无限不循环小数。Where y is the second dimensionality reduction data set, and e is an infinite non-cyclic decimal.
本申请较佳实施例所述损失值T为：The loss value T in the preferred embodiment of the present application is:
Figure PCTCN2019117352-appb-000010
Figure PCTCN2019117352-appb-000010
其中，n为所述训练集的大小，y t为所述训练值，μ t为所述标签集。 Wherein, n is the size of the training set, y t is the training value, and μ t is the label set.
S3、接收输入的声音数据并输入至所述人声检测模型，所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。S3. Receive the input sound data and input it to the human voice detection model. The human voice detection model judges whether the sound data includes human voice and outputs a judgment result.
发明还提供一种人声智能检测装置。参照图2所示，为本申请一实施例提供的人声智能检测装置的内部结构示意图。The invention also provides a human voice intelligent detection device. Referring to FIG. 2, it is a schematic diagram of the internal structure of a human voice intelligent detection device provided by an embodiment of this application.
在本实施例中，所述人声智能检测装置1可以是PC(Personal Computer，个人电脑)，或者是智能手机、平板电脑、便携计算机等终端设备，也可以是一种服务器等。该人声智能检测装置1至少包括存储器11、处理器12，通信总线13，以及网络接口14。In this embodiment, the human voice intelligent detection device 1 may be a PC (Personal Computer, personal computer), or a terminal device such as a smart phone, a tablet computer, or a portable computer, or a server. The human voice intelligent detection device 1 at least includes a memory 11, a processor 12, a communication bus 13, and a network interface 14.
其中，存储器11至少包括一种类型的可读存储介质，所述可读存储介质包括闪存、硬盘、多媒体卡、卡型存储器(例如，SD或DX存储器等)、磁性存储器、磁盘、光盘等。存储器11在一些实施例中可以是人声智能检测装置1的内部存储单元，例如该人声智能检测装置1的硬盘。存储器11在另一些实施例中也可以是人声智能检测装置1的外部存储设备，例如人声智能检测装置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。进一步地，存储器11还可以既包括人声智能检测装置1的内部存储单元也包括外部存储设备。存储器11不仅可以用于存储安装于人声智能检测装置1的应用软件及各 类数据，例如人声智能检测程序01的代码等，还可以用于暂时地存储已经输出或者将要输出的数据。Wherein, the memory 11 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, etc. The memory 11 may be an internal storage unit of the human voice intelligent detection device 1 in some embodiments, such as a hard disk of the human voice intelligent detection device 1. In other embodiments, the memory 11 may also be an external storage device of the human voice intelligent detection device 1, for example, a plug-in hard disk equipped on the human voice intelligent detection device 1, a smart media card (SMC), and a secure digital (Secure Digital, SD) card, flash card (Flash Card), etc. Further, the memory 11 may also include both an internal storage unit of the human voice intelligent detection device 1 and an external storage device. The memory 11 can be used not only to store application software and various data installed in the human voice intelligent detection device 1, such as the code of the human voice intelligent detection program 01, etc., but also to temporarily store data that has been output or will be output.
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit，CPU)、控制器、微控制器、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，例如执行人声智能检测程序01(所述人声智能检测程序01实质是一个软件系统)等。In some embodiments, the processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip, and is used to run the program code or processing stored in the memory 11 Data, for example, execute the human voice intelligent detection program 01 (the human voice intelligent detection program 01 is essentially a software system) and so on.
通信总线13用于实现这些组件之间的连接通信。The communication bus 13 is used to realize the connection and communication between these components.
网络接口14可选的可以包括标准的有线接口、无线接口(如WI-FI接口)，通常用于在该装置1与其他电子设备之间建立通信连接。The network interface 14 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 device 1 and other electronic devices.
可选地，该装置1还可以包括用户接口，用户接口可以包括显示器(Display)、输入单元比如键盘(Keyboard)，可选的用户接口还可以包括标准的有线接口、无线接口。可选地，在一些实施例中，显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。其中，显示器也可以适当的称为显示屏或显示单元，用于显示在人声智能检测装置1中处理的信息以及用于显示可视化的用户界面。Optionally, the device 1 may also include a user interface. The user interface may include a display (Display) and an input unit such as a keyboard (Keyboard). The optional user interface may also include a standard wired interface and a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light emitting diode) touch device, etc. Among them, the display may also be appropriately called a display screen or a display unit, which is used to display the information processed in the human voice intelligent detection device 1 and to display a visualized user interface.
图2仅示出了具有组件11-14以及人声智能检测程序01的人声智能检测装置1，本领域技术人员可以理解的是，图1示出的结构并不构成对人声智能检测装置1的限定，可以包括比图示更少或者更多的部件，或者组合某些部件，或者不同的部件布置。Figure 2 only shows the human voice intelligent detection device 1 with components 11-14 and human voice intelligent detection program 01. Those skilled in the art will understand that the structure shown in Figure 1 does not constitute a human voice intelligent detection device The definition of 1 may include fewer or more components than shown, or a combination of certain components, or different component arrangements.
在图2所示的装置1实施例中，存储器11中存储有人声智能检测程序01；处理器12执行存储器11中存储的人声智能检测程序01时实现如下步骤：In the embodiment of the device 1 shown in FIG. 2, the human voice intelligent detection program 01 is stored in the memory 11; when the processor 12 executes the human voice intelligent detection program 01 stored in the memory 11, the following steps are implemented:
步骤一、数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数。 Step 1. The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set A pre-processing operation including pre-emphasis and windowing and framing is performed, the training set completed by the pre-processing operation is input to the human voice detection model, and the label set is input to the loss function.
本申请较佳实施例，所述包括人声数据的正样本集是在安静的环境下通过麦克风录取，所述麦克风录取的采样频率16kHz、采样大小16bits，且参与录取的人员至少录制两段不同人声数据，一段使用标准普通话录取，另一段使用所述录取人员的地方方言录取。所述正样本集内每段人声数据时长不少 于10秒。In a preferred embodiment of the present application, the positive sample set including human voice data is recorded by a microphone in a quiet environment, the sampling frequency of the microphone recording is 16kHz, the sampling size is 16bits, and the persons participating in the admission record at least two different sections The vocal data, one section is admitted in standard Mandarin, and the other section is admitted in the local dialect of the admitted person. The duration of each segment of human voice data in the positive sample set is not less than 10 seconds.
本申请较佳实施例，所述负样本集来源于音频数据集AudioSet中包括多条人工标记的声音剪辑片段，所述AudioSet是目前开放的大规模且完善的音频数据集，进一步地，所述多条人工标记的声音剪辑片段包括2084320条人工标记的每段10秒长度的声音剪辑片段。In a preferred embodiment of the present application, the negative sample set comes from the audio data set AudioSet, which includes multiple manually marked sound clips. The AudioSet is a large-scale and complete audio data set that is currently open. Further, the Multiple manually marked sound clips include 2084320 manually marked sound clips each with a length of 10 seconds.
本申请较佳实施所述预加重预处理操作是提高所述训练集的高频音域部分，使所述训练集的音域低频到音域高频的信号频谱变得平坦，同时还能抑制随机噪声和直流漂移的影响，进一步地，所述预加重是基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重即所述预加重的方法为：The preferred implementation of the pre-emphasis pre-processing operation in this application is to improve the high-frequency range part of the training set, so that the signal spectrum from the low-frequency range to the high-frequency range of the training set becomes flat, and at the same time, it can suppress random noise and Further, the pre-emphasis is based on the digital filter to pre-emphasize the sound frequency of the training set, and the method of the pre-emphasis, that is, the pre-emphasis, is:
H(z)＝1-μz -1 H(z) = 1-μz -1
其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
本申请较佳实施所述加窗分帧是根据在小范围的时间内，所述训练集的音频信号保持不变的特点，对所述训练集的音频信号进行分帧处理，进一地，所述加窗分帧基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：The preferred implementation of the windowing and framing in this application is based on the feature that the audio signal of the training set remains unchanged within a small range of time, and the audio signal of the training set is subjected to framing processing, and further, The windowing and framing is based on the pre-emphasized training set, and the windowing and framing processing is performed according to the Hamming window method, and the Hamming window method ω(n) is:
Figure PCTCN2019117352-appb-000011
Figure PCTCN2019117352-appb-000011
其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
步骤二、所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练。Step 2: The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the The training value is calculated to obtain a loss value, and the size of the loss value and a preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits the training.
本申请较佳实施例所述人声检测模型接收所述预处理操作完成的训练集，将所述训练集输入至第一层卷积层，所述第一层卷积层进行卷积操作后得到卷积数据集输入至第一层池化层；其后，所述第一层池化层对所述卷积数据集进行最大化池化操作后得到降维数据集输入至第二层卷积层，所述第二层卷积层进行所述卷积操作后输入至第二层池化层进行所述最大化池化操作，直到最终输入至全连接层；所述全连接层结合激活函数计算得到所述训练值；The human voice detection model described in the preferred embodiment of the present application receives the training set completed by the preprocessing operation, and inputs the training set to the first layer of convolutional layer. After the first layer of convolutional layer performs convolution operation Obtain the convolutional data set and input it to the first-layer pooling layer; then, the first-layer pooling layer performs a maximization pooling operation on the convolutional data set to obtain a dimensionality reduction data set and input it to the second-layer volume Layers, the second convolutional layer performs the convolution operation and then inputs to the second pooling layer for the maximization pooling operation, until finally input to the fully connected layer; the fully connected layer is combined with activation Function calculation to obtain the training value;
本申请较佳实施例所述卷积操作为：The convolution operation described in the preferred embodiment of this application is:
Figure PCTCN2019117352-appb-000012
Figure PCTCN2019117352-appb-000012
其中ω’为输出数据，ω为输入数据，k为卷积核的大小，s为卷积操作的步幅，p为数据补零矩阵；Where ω’ is the output data, ω is the input data, k is the size of the convolution kernel, s is the step size of the convolution operation, and p is the data zero-filling matrix;
本申请较佳实施例所述激活函数为：The activation function described in the preferred embodiment of this application is:
Figure PCTCN2019117352-appb-000013
Figure PCTCN2019117352-appb-000013
其中y为所述第二降维数据集，e为无限不循环小数。Where y is the second dimensionality reduction data set, and e is an infinite non-cyclic decimal.
本申请较佳实施例所述损失值T为：The loss value T in the preferred embodiment of the present application is:
Figure PCTCN2019117352-appb-000014
Figure PCTCN2019117352-appb-000014
其中，n为所述训练集的大小，y t为所述训练值，μ t为所述标签集。 Wherein, n is the size of the training set, y t is the training value, and μ t is the label set.
步骤三、接收输入的声音数据并输入至所述人声检测模型，所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。Step 3: The input voice data is received and input to the human voice detection model, and the human voice detection model judges whether the voice data includes human voice and outputs the judgment result.
可选地，在其他实施例中，人声智能检测程序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，并由一个或多个处理器(本实施例为处理器12)所执行以完成本申请，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，用于描述人声智能检测程序在人声智能检测装置中的执行过程。Optionally, in other embodiments, the human voice intelligent detection program can also be divided into one or more modules, and the one or more modules are stored in the memory 11 and run by one or more processors (this embodiment It is 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, and is used to describe the execution process of the human voice intelligent detection program in the human voice intelligent detection device .
例如，参照图3所示，为本申请人声智能检测装置一实施例中的人声智能检测程序的程序模块示意图，该实施例中，所述人声智能检测程序可以被分割为数据接收模块10、模型训练模块20、人声结果输出模块30，示例性地：For example, referring to FIG. 3, it is a schematic diagram of the program modules of the human voice intelligent detection program in an embodiment of the applicant's voice intelligent detection device. In this embodiment, the human voice intelligent detection program can be divided into data receiving modules 10. Model training module 20, human voice result output module 30, exemplarily:
所述数据接收模块10用于：接收包括人声数据的正样本集、不包括人声数据的负样本集和标签集，所述正样本集和所述负样本集统称训练集，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数。The data receiving module 10 is configured to receive a positive sample set including human voice data, a negative sample set not including human voice data, and a label set. The positive sample set and the negative sample set are collectively referred to as a training set, and the The training set is subjected to pre-processing operations including pre-emphasis and windowing and framing, the training set completed by the pre-processing operation is input to the human voice detection model, and the label set is input to the loss function.
所述模型训练模块20用于：所述人声检测模型接收所述预处理操作完成的训练集进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练。The model training module 20 is configured to: the human voice detection model receives the training set completed by the preprocessing operation to obtain training values, and inputs the training values into the loss function, which is based on the The label set and the training value are calculated to obtain a loss value, and the size of the loss value and a preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training.
所述人声结果输出模块30用于：接收输入的声音数据并输入至所述人声检测模型，所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The human voice result output module 30 is configured to receive input voice data and input it to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
上述数据接收模块10、模型训练模块20、人声结果输出模块30等程序模块被执行时所实现的功能或操作步骤与上述实施例大体相同，在此不再赘述。The functions or operation steps implemented by the program modules such as the data receiving module 10, the model training module 20, and the human voice result output module 30 when executed are substantially the same as those in the foregoing embodiment, and will not be repeated here.
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上存储有人声智能检测程序，所述人声智能检测程序可被一个或多个处理器执行，以实现如下操作：In addition, the embodiment of the present application also proposes a computer-readable storage medium. The computer-readable storage medium stores a human voice intelligent detection program, and the human voice intelligent detection program can be executed by one or more processors to realize Do as follows:
接收包括人声数据的正样本集、不包括人声数据的负样本集和标签集，所述正样本集和所述负样本集统称训练集，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数。Receive a positive sample set including human voice data, a negative sample set that does not include human voice data, and a label set, the positive sample set and the negative sample set are collectively referred to as a training set, and pre-emphasis and windowing are performed on the training set The pre-processing operation of framing is to input the training set completed by the pre-processing operation into the human voice detection model, and input the label set into the loss function.
所述人声检测模型接收所述预处理操作完成的训练集进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练。The human voice detection model receives the training set completed by the preprocessing operation for training to obtain training values, and inputs the training values to the loss function, and the loss function is calculated based on the label set and the training value The loss value is obtained, and the size of the loss value and a preset threshold is judged, and the human voice detection model exits training when the loss value is less than the preset threshold.
接收输入的声音数据并输入至所述人声检测模型，所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The input voice data is received and input to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
本申请计算机可读存储介质具体实施方式与上述人声智能检测装置和方法各实施例基本相同，在此不作累述。The specific implementation of the computer-readable storage medium of the present application is basically the same as the foregoing embodiments of the human voice intelligent detection device and method, and will not be repeated here.
需要说明的是，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。并且本文中的术语“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。It should be noted that the serial numbers of the above embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. And the terms "include", "include" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements not only includes those elements, but also includes The other elements listed may also include elements inherent to the process, device, article, or method. Without more restrictions, the element defined by the sentence "including a..." does not exclude the existence of other identical elements in the process, device, article, or method that includes the element.
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设备(可以是手机，计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。Through the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiments can be implemented by means of software plus the necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is better.的实施方式。 Based on this understanding, the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM/RAM) as described above. , Magnetic disk, optical disk), including several instructions to make a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) execute the method described in each embodiment of the present application.
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。The above are only preferred embodiments of this application, and do not limit the scope of this application. Any equivalent structure or equivalent process transformation made using the content of the description and drawings of this application, or directly or indirectly used in other related technical fields , The same reason is included in the scope of patent protection of this application.

Claims (20)

  1. 一种人声智能检测方法，其特征在于，所述方法包括：A human voice intelligent detection method, characterized in that the method includes:
    数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数；The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set includes pre-processing. Preprocessing operations of emphasis and windowing and framing, input the training set completed by the preprocessing operation to the human voice detection model, and input the label set to the loss function;
    所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练；The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the training value The loss value is calculated, and the size of the loss value and the preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
    接收输入的声音数据并输入至所述人声检测模型，利用所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The input voice data is received and input to the human voice detection model, and the human voice detection model is used to determine whether the voice data includes human voice and output the judgment result.
  2. 如权利要求1所述的人声智能检测方法，其特征在于，所述数据处理层接收包括正样本集和负样本集的训练集和标签集，包括：The human voice intelligent detection method according to claim 1, wherein the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, including:
    提取预设音频数据集AudioSet中包括的多条人工标记的声音剪辑片段做为所述负样本集；Extracting a plurality of manually marked sound clips included in the preset audio data set AudioSet as the negative sample set;
    录制多种采样频率的人声，构建所述正样本集；Recording vocals of multiple sampling frequencies to construct the positive sample set;
    基于所述正样本集和所述负样本集建立对应的标签集。A corresponding label set is established based on the positive sample set and the negative sample set.
  3. 如权利要求2所述的人声智能检测方法，其特征在于，对所述训练集进行包括预加重和加窗分帧的预处理操作，包括：The human voice intelligent detection method according to claim 2, wherein performing pre-processing operations including pre-emphasis and windowing and framing on the training set includes:
    基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
    H(z)＝1-μz -1 H(z) = 1-μz -1
    其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
    基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
    Figure PCTCN2019117352-appb-100001
    Figure PCTCN2019117352-appb-100001
    其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
  4. 如权利要求1至3中任意一项所述的人声智能检测方法，其特征在于，所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，包括：The human voice intelligent detection method according to any one of claims 1 to 3, wherein the human voice detection model receives a training set completed by the preprocessing operation and performs training to obtain a training value, comprising:
    将所述训练集输入至所述人声检测模型的第一层卷积层进行卷积操作，得到第一卷积数据集，并将所述第一卷积数据集输入至第一层池化层；Input the training set to the first convolutional layer of the human voice detection model to perform a convolution operation to obtain a first convolutional data set, and input the first convolutional data set to the first layer of pooling Floor;
    所述第一层池化层对所述第一卷积数据集进行最大化池化操作，得到第一降维数据集，并将所述第一降维数据集输入至第二层卷积层进行所述卷积操作，得到第二卷积数据集，将所述第二卷积数据集输入至第二层池化层进行所述最大化池化操作，得到第二降维数据集，并将所述第二降维数据集输入至全连接层；The first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
    所述全连接层结合激活函数对所述第二降维数据集执行计算，得到所述训练值。The fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
  5. 如权利要求4所述的人声智能检测方法，其特征在于，所述卷积操作为：The human voice intelligent detection method according to claim 4, wherein the convolution operation is:
    Figure PCTCN2019117352-appb-100002
    Figure PCTCN2019117352-appb-100002
    其中ω’为输出数据，ω为输入数据，k为卷积核的大小，s为所述卷积操作的步幅，p为数据补零矩阵；Where ω'is output data, ω is input data, k is the size of the convolution kernel, s is the step size of the convolution operation, and p is the data zero-filling matrix;
    所述激活函数为：The activation function is:
    Figure PCTCN2019117352-appb-100003
    Figure PCTCN2019117352-appb-100003
    其中y为所述第二降维数据集，e为无限不循环小数。Where y is the second dimensionality reduction data set, and e is an infinite non-cyclic decimal.
  6. 一种人声智能检测装置，其特征在于，所述装置包括存储器和处理器，所述存储器上存储有可在所述处理器上运行的人声智能检测程序，所述人声智能检测程序被所述处理器执行时实现如下步骤：A human voice intelligent detection device, characterized in that the device includes a memory and a processor, the memory stores a human voice intelligent detection program that can run on the processor, and the human voice intelligent detection program is The processor implements the following steps when executing:
    数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数；The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set includes pre-processing. Preprocessing operations of emphasis and windowing and framing, input the training set completed by the preprocessing operation to the human voice detection model, and input the label set to the loss function;
    所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所 述损失值小于所述预设阈值时，所述人声检测模型退出训练；The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the training value The loss value is calculated, and the size of the loss value and the preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
    接收输入的声音数据并输入至所述人声检测模型，所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The input voice data is received and input to the human voice detection model, and the human voice detection model determines whether the voice data includes human voice and outputs the judgment result.
  7. 如权利要求6所述的人声智能检测装置，其特征在于，所述数据处理层接收包括正样本集和负样本集的训练集和标签集，包括：7. The human voice intelligent detection device according to claim 6, wherein the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, including:
    提取预设音频数据集AudioSet中包括的多条人工标记的声音剪辑片段做为所述负样本集；Extracting a plurality of manually marked sound clips included in the preset audio data set AudioSet as the negative sample set;
    录制多种采样频率的人声，构建所述正样本集；Recording vocals of multiple sampling frequencies to construct the positive sample set;
    基于所述正样本集和所述负样本集建立对应的标签集。A corresponding label set is established based on the positive sample set and the negative sample set.
  8. 如权利要求7所述的人声智能检测装置，其特征在于，对所述训练集进行包括预加重和加窗分帧的预处理操作，包括：8. The human voice intelligent detection device according to claim 7, wherein the pre-processing operation including pre-emphasis and windowing and framing on the training set comprises:
    基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
    H(z)＝1-μz -1 H(z) = 1-μz -1
    其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
    基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
    Figure PCTCN2019117352-appb-100004
    Figure PCTCN2019117352-appb-100004
    其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
  9. 如权利要求6至8任意一项所述的人声智能检测装置，其特征在于，所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，包括：8. The human voice intelligent detection device according to any one of claims 6 to 8, wherein the human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, comprising:
    将所述训练集输入至所述人声检测模型的第一层卷积层进行卷积操作，得到第一卷积数据集，并将所述第一卷积数据集输入至第一层池化层；Input the training set to the first convolutional layer of the human voice detection model to perform a convolution operation to obtain a first convolutional data set, and input the first convolutional data set to the first layer of pooling Floor;
    所述第一层池化层对所述第一卷积数据集进行最大化池化操作，得到第一降维数据集，并将所述第一降维数据集输入至第二层卷积层进行所述卷积操作，得到第二卷积数据集，将所述第二卷积数据集输入至第二层池化层进行所述最大化池化操作，得到第二降维数据集，并将所述第二降维数据集输入至全连接层；The first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
    所述全连接层结合激活函数对所述第二降维数据集执行计算，得到所述训练值。The fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
  10. 如权利要求9所述的人声智能检测装置，其特征在于，所述卷积操作为：The human voice intelligent detection device according to claim 9, wherein the convolution operation is:
    Figure PCTCN2019117352-appb-100005
    Figure PCTCN2019117352-appb-100005
    其中ω’为输出数据，ω为输入数据，k为卷积核的大小，s为所述卷积操作的步幅，p为数据补零矩阵；Where ω'is output data, ω is input data, k is the size of the convolution kernel, s is the step size of the convolution operation, and p is the data zero-filling matrix;
    所述激活函数为：The activation function is:
    Figure PCTCN2019117352-appb-100006
    Figure PCTCN2019117352-appb-100006
    其中y为所述第二降维数据集，e为无限不循环小数。Where y is the second dimensionality reduction data set, and e is an infinite non-cyclic decimal.
  11. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有人声智能检测程序，所述人声智能检测程序可被一个或者多个处理器执行，以实现如下步骤：A computer-readable storage medium, characterized in that a human voice intelligent detection program is stored on the computer-readable storage medium, and the human voice intelligent detection program can be executed by one or more processors to implement the following steps:
    数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数；The data processing layer receives a training set and a label set including a positive sample set and a negative sample set, where the positive sample set includes human voice data and the negative sample set does not include human voice data, and the training set includes pre-processing. Preprocessing operations of emphasis and windowing and framing, input the training set completed by the preprocessing operation to the human voice detection model, and input the label set to the loss function;
    所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练；The human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the label set and the training value The loss value is calculated, and the size of the loss value and the preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
    接收输入的声音数据并输入至所述人声检测模型，利用所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The input voice data is received and input to the human voice detection model, and the human voice detection model is used to determine whether the voice data includes human voice and output the judgment result.
  12. 如权利要求11所述的计算机可读存储介质，其特征在于，所述数据处理层接收包括正样本集和负样本集的训练集和标签集，包括：11. The computer-readable storage medium of claim 11, wherein the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, comprising:
    提取预设音频数据集AudioSet中包括的多条人工标记的声音剪辑片段做为所述负样本集；Extracting a plurality of manually marked sound clips included in the preset audio data set AudioSet as the negative sample set;
    录制多种采样频率的人声，构建所述正样本集；Recording vocals of multiple sampling frequencies to construct the positive sample set;
    基于所述正样本集和所述负样本集建立对应的标签集。A corresponding label set is established based on the positive sample set and the negative sample set.
  13. 如权利要求12所述的计算机可读存储介质，其特征在于，对所述训 练集进行包括预加重和加窗分帧的预处理操作，包括：The computer-readable storage medium of claim 12, wherein performing pre-processing operations including pre-emphasis and windowing and framing on the training set comprises:
    基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
    H(z)＝1-μz -1 H(z) = 1-μz -1
    其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
    基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
    Figure PCTCN2019117352-appb-100007
    Figure PCTCN2019117352-appb-100007
    其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
  14. 如权利要求11至13中任意一项所述的计算机可读存储介质，其特征在于，所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，包括：The computer-readable storage medium according to any one of claims 11 to 13, wherein the human voice detection model receives a training set completed by the preprocessing operation and performs training to obtain a training value, comprising:
    将所述训练集输入至所述人声检测模型的第一层卷积层进行卷积操作，得到第一卷积数据集，并将所述第一卷积数据集输入至第一层池化层；Input the training set to the first convolutional layer of the human voice detection model to perform a convolution operation to obtain a first convolutional data set, and input the first convolutional data set to the first layer of pooling Floor;
    所述第一层池化层对所述第一卷积数据集进行最大化池化操作，得到第一降维数据集，并将所述第一降维数据集输入至第二层卷积层进行所述卷积操作，得到第二卷积数据集，将所述第二卷积数据集输入至第二层池化层进行所述最大化池化操作，得到第二降维数据集，并将所述第二降维数据集输入至全连接层；The first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
    所述全连接层结合激活函数对所述第二降维数据集执行计算，得到所述训练值。The fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
  15. 如权利要求14所述的计算机可读存储介质，其特征在于，所述卷积操作为：The computer-readable storage medium of claim 14, wherein the convolution operation is:
    Figure PCTCN2019117352-appb-100008
    Figure PCTCN2019117352-appb-100008
    其中ω’为输出数据，ω为输入数据，k为卷积核的大小，s为所述卷积操作的步幅，p为数据补零矩阵；Where ω'is output data, ω is input data, k is the size of the convolution kernel, s is the step size of the convolution operation, and p is the data zero-filling matrix;
    所述激活函数为：The activation function is:
    Figure PCTCN2019117352-appb-100009
    Figure PCTCN2019117352-appb-100009
    其中y为所述第二降维数据集，e为无限不循环小数。Where y is the second dimensionality reduction data set, and e is an infinite non-cyclic decimal.
  16. 一种人声智能检测系统，其特征在于，所述人声智能检测系统包括：A human voice intelligent detection system, characterized in that the human voice intelligent detection system includes:
    数据接收模块，用于：数据处理层接收包括正样本集和负样本集的训练集和标签集，其中，所述正样本集包括人声数据以及所述负样本集不包括人声数据，对所述训练集进行包括预加重和加窗分帧的预处理操作，将所述预处理操作完成的训练集输入至人声检测模型，将所述标签集输入至损失函数；The data receiving module is configured to: the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, wherein the positive sample set includes human voice data and the negative sample set does not include human voice data, right The training set is subjected to pre-processing operations including pre-emphasis and windowing and framing, the training set completed by the pre-processing operation is input to a human voice detection model, and the label set is input to a loss function;
    模型训练模块，用于：所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，并将所述训练值输入至所述损失函数，所述损失函数基于所述标签集和所述训练值计算得到损失值，判断所述损失值与预设阈值的大小，直至所述损失值小于所述预设阈值时，所述人声检测模型退出训练；The model training module is configured to: the human voice detection model receives the training set completed by the preprocessing operation and performs training to obtain training values, and inputs the training values to the loss function, which is based on the The label set and the training value are calculated to obtain a loss value, and the size of the loss value and a preset threshold is judged, until the loss value is less than the preset threshold, the human voice detection model exits training;
    人声结果输出模块，用于：接收输入的声音数据并输入至所述人声检测模型，利用所述人声检测模型判断所述声音数据是否包括人声并输出判断结果。The human voice result output module is configured to: receive the input voice data and input it into the human voice detection model, use the human voice detection model to determine whether the voice data includes human voice and output the judgment result.
  17. 如权利要求16所述的人声智能检测系统，其特征在于，所述数据处理层接收包括正样本集和负样本集的训练集和标签集，包括：The human voice intelligent detection system according to claim 16, wherein the data processing layer receives a training set and a label set including a positive sample set and a negative sample set, including:
    提取预设音频数据集AudioSet中包括的多条人工标记的声音剪辑片段做为所述负样本集；Extracting a plurality of manually marked sound clips included in the preset audio data set AudioSet as the negative sample set;
    录制多种采样频率的人声，构建所述正样本集；Recording vocals of multiple sampling frequencies to construct the positive sample set;
    基于所述正样本集和所述负样本集建立对应的标签集。A corresponding label set is established based on the positive sample set and the negative sample set.
  18. 如权利要求17所述的人声智能检测系统，其特征在于，对所述训练集进行包括预加重和加窗分帧的预处理操作，包括：The human voice intelligent detection system according to claim 17, wherein the pre-processing operations including pre-emphasis and windowing and framing on the training set include:
    基于数字滤波器对所述训练集的声音频率进行预加重，所述预加重的方法为：Perform pre-emphasis on the sound frequency of the training set based on a digital filter, and the pre-emphasis method is:
    H(z)＝1-μz -1 H(z) = 1-μz -1
    其中，H(z)为所述预加重后的训练集，z为所述声音频率，μ为预加重系数；Wherein, H(z) is the training set after the pre-emphasis, z is the sound frequency, and μ is the pre-emphasis coefficient;
    基于所述预加重后的训练集，根据汉明窗法进行加窗分帧处理，所述汉明窗法ω(n)为：Based on the pre-emphasized training set, perform windowing and framing processing according to the Hamming window method, and the Hamming window method ω(n) is:
    Figure PCTCN2019117352-appb-100010
    Figure PCTCN2019117352-appb-100010
    其中，n为所述预加重后的训练集，N为所述汉明窗法的窗长，cos为余弦 函数。Wherein, n is the training set after the pre-emphasis, N is the window length of the Hamming window method, and cos is the cosine function.
  19. 如权利要求16至18所述的人声智能检测系统，其特征在于，所述人声检测模型接收所述预处理操作完成的训练集并进行训练得到训练值，包括：The human voice intelligent detection system according to claims 16 to 18, wherein the human voice detection model receives the training set completed by the pre-processing operation and performs training to obtain training values, comprising:
    将所述训练集输入至所述人声检测模型的第一层卷积层进行卷积操作，得到第一卷积数据集，并将所述第一卷积数据集输入至第一层池化层；Input the training set to the first convolutional layer of the human voice detection model to perform a convolution operation to obtain a first convolutional data set, and input the first convolutional data set to the first layer of pooling Floor;
    所述第一层池化层对所述第一卷积数据集进行最大化池化操作，得到第一降维数据集，并将所述第一降维数据集输入至第二层卷积层进行所述卷积操作，得到第二卷积数据集，将所述第二卷积数据集输入至第二层池化层进行所述最大化池化操作，得到第二降维数据集，并将所述第二降维数据集输入至全连接层；The first-level pooling layer performs a maximization pooling operation on the first convolutional data set to obtain a first dimensionality reduction data set, and inputs the first dimensionality reduction data set to the second-level convolutional layer Perform the convolution operation to obtain a second convolution data set, and input the second convolution data set to the second pooling layer to perform the maximization pooling operation to obtain a second dimensionality reduction data set, and Input the second dimensionality reduction data set to the fully connected layer;
    所述全连接层结合激活函数对所述第二降维数据集执行计算，得到所述训练值。The fully connected layer combines an activation function to perform calculation on the second dimensionality reduction data set to obtain the training value.
  20. 如权利要求19所述的人声智能检测系统，其特征在于，所述卷积操作为：The human voice intelligent detection system according to claim 19, wherein the convolution operation is:
    Figure PCTCN2019117352-appb-100011
    Figure PCTCN2019117352-appb-100011
    其中ω’为输出数据，ω为输入数据，k为卷积核的大小，s为所述卷积操作的步幅，p为数据补零矩阵；Where ω'is output data, ω is input data, k is the size of the convolution kernel, s is the step size of the convolution operation, and p is the data zero-filling matrix;
    所述激活函数为：The activation function is:
    Figure PCTCN2019117352-appb-100012
    Figure PCTCN2019117352-appb-100012
    其中y为所述第二降维数据集，e为无限不循环小数。Where y is the second dimensionality reduction data set, and e is an infinite non-cyclic decimal.
PCT/CN2019/117352 2019-05-29 2019-11-12 Human voice smart detection method and apparatus, and computer readable storage medium WO2020238046A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910468133.4 2019-05-29
CN201910468133.4A CN110246506A (en) 2019-05-29 2019-05-29 Voice intelligent detecting method, device and computer readable storage medium

Publications (1)

Publication Number Publication Date
WO2020238046A1 true WO2020238046A1 (en) 2020-12-03

Family

ID=67885602

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/117352 WO2020238046A1 (en) 2019-05-29 2019-11-12 Human voice smart detection method and apparatus, and computer readable storage medium

Country Status (2)

Country Link
CN (1) CN110246506A (en)
WO (1) WO2020238046A1 (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110246506A (en) * 2019-05-29 2019-09-17 平安科技（深圳）有限公司 Voice intelligent detecting method, device and computer readable storage medium
CN110751941B (en) * 2019-09-18 2023-05-26 平安科技（深圳）有限公司 Speech synthesis model generation method, device, equipment and storage medium
CN110765868A (en) * 2019-09-18 2020-02-07 平安科技（深圳）有限公司 Lip reading model generation method, device, equipment and storage medium
CN111221942A (en) * 2020-01-09 2020-06-02 平安科技（深圳）有限公司 Intelligent text conversation generation method and device and computer readable storage medium
CN111243609B (en) * 2020-01-10 2023-07-14 平安科技（深圳）有限公司 Method and device for intelligent detection of effective voice and computer readable storage medium
CN113936694B (en) * 2021-12-17 2022-03-18 珠海普林芯驰科技有限公司 Real-time human voice detection method, computer device and computer readable storage medium

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2016003299A1 (en) * 2014-07-04 2016-01-07 Intel Corporation Replay attack detection in automatic speaker verification systems
CN105374357A (en) * 2015-11-23 2016-03-02 青岛海尔智能技术研发有限公司 Voice recognition method, device and voice control system
CN108806698A (en) * 2018-03-15 2018-11-13 中山大学 A kind of camouflage audio recognition method based on convolutional neural networks
CN108986824A (en) * 2018-07-09 2018-12-11 宁波大学 A kind of voice playback detection method
CN109350032A (en) * 2018-10-16 2019-02-19 武汉中旗生物医疗电子有限公司 A kind of classification method, system, electronic equipment and storage medium
CN109599117A (en) * 2018-11-14 2019-04-09 厦门快商通信息技术有限公司 A kind of audio data recognition methods and human voice anti-replay identifying system
CN110246506A (en) * 2019-05-29 2019-09-17 平安科技（深圳）有限公司 Voice intelligent detecting method, device and computer readable storage medium

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5737808B2 (en) * 2011-08-31 2015-06-17 日本放送協会 Sound processing apparatus and program thereof
CN107086036A (en) * 2017-04-19 2017-08-22 杭州派尼澳电子科技有限公司 A kind of freeway tunnel method for safety monitoring
CN107393542B (en) * 2017-06-28 2020-05-19 北京林业大学 Bird species identification method based on two-channel neural network
CN108665005B (en) * 2018-05-16 2021-12-07 南京信息工程大学 Method for improving CNN-based image recognition performance by using DCGAN
CN108922561A (en) * 2018-06-04 2018-11-30 平安科技（深圳）有限公司 Speech differentiation method, apparatus, computer equipment and storage medium
CN109166593B (en) * 2018-08-17 2021-03-16 腾讯音乐娱乐科技（深圳）有限公司 Audio data processing method, device and storage medium
CN109754812A (en) * 2019-01-30 2019-05-14 华南理工大学 A kind of voiceprint authentication method of the anti-recording attack detecting based on convolutional neural networks

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2016003299A1 (en) * 2014-07-04 2016-01-07 Intel Corporation Replay attack detection in automatic speaker verification systems
CN105374357A (en) * 2015-11-23 2016-03-02 青岛海尔智能技术研发有限公司 Voice recognition method, device and voice control system
CN108806698A (en) * 2018-03-15 2018-11-13 中山大学 A kind of camouflage audio recognition method based on convolutional neural networks
CN108986824A (en) * 2018-07-09 2018-12-11 宁波大学 A kind of voice playback detection method
CN109350032A (en) * 2018-10-16 2019-02-19 武汉中旗生物医疗电子有限公司 A kind of classification method, system, electronic equipment and storage medium
CN109599117A (en) * 2018-11-14 2019-04-09 厦门快商通信息技术有限公司 A kind of audio data recognition methods and human voice anti-replay identifying system
CN110246506A (en) * 2019-05-29 2019-09-17 平安科技（深圳）有限公司 Voice intelligent detecting method, device and computer readable storage medium

Also Published As

Publication number Publication date
CN110246506A (en) 2019-09-17

Similar Documents

Publication Publication Date Title
WO2020238046A1 (en) Human voice smart detection method and apparatus, and computer readable storage medium
WO2021208287A1 (en) Voice activity detection method and apparatus for emotion recognition, electronic device, and storage medium
CN108564954B (en) Deep neural network model, electronic device, identity verification method, and storage medium
WO2020237855A1 (en) Sound separation method and apparatus, and computer readable storage medium
WO2020232860A1 (en) Speech synthesis method and apparatus, and computer readable storage medium
CN109859772B (en) Emotion recognition method, emotion recognition device and computer-readable storage medium
WO2019085329A1 (en) Recurrent neural network-based personal character analysis method, device, and storage medium
WO2021004481A1 (en) Media files recommending method and device
WO2020238045A1 (en) Intelligent speech recognition method and apparatus, and computer-readable storage medium
WO2019179029A1 (en) Electronic device, identity verification method and computer-readable storage medium
CN112562691A (en) Voiceprint recognition method and device, computer equipment and storage medium
CN112418059B (en) Emotion recognition method and device, computer equipment and storage medium
WO2021051514A1 (en) Speech identification method and apparatus, computer device and non-volatile storage medium
WO2020098523A1 (en) Voice recognition method and device and computing device
WO2021196475A1 (en) Intelligent language fluency recognition method and apparatus, computer device, and storage medium
CN113129867B (en) Training method of voice recognition model, voice recognition method, device and equipment
CN116340778B (en) Medical large model construction method based on multiple modes and related equipment thereof
WO2021027117A1 (en) Speech emotion recognition method and appartus, and computer-readable storage medium
CN110136726A (en) A kind of estimation method, device, system and the storage medium of voice gender
Huijuan et al. Coarse-to-fine speech emotion recognition based on multi-task learning
CN112489628B (en) Voice data selection method and device, electronic equipment and storage medium
CN114138960A (en) User intention identification method, device, equipment and medium
WO2019228140A1 (en) Instruction execution method and apparatus, storage medium, and electronic device
WO2021196477A1 (en) Risk user identification method and apparatus based on voiceprint characteristics and associated graph data
CN110415708A (en) Method for identifying speaker, device, equipment and storage medium neural network based

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: 19931235

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: 19931235

Country of ref document: EP

Kind code of ref document: A1