WO2020000523A1 - 一种信号处理方法及装置 - Google Patents

一种信号处理方法及装置 Download PDF

Info

Publication number
WO2020000523A1
WO2020000523A1 PCT/CN2018/095377 CN2018095377W WO2020000523A1 WO 2020000523 A1 WO2020000523 A1 WO 2020000523A1 CN 2018095377 W CN2018095377 W CN 2018095377W WO 2020000523 A1 WO2020000523 A1 WO 2020000523A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
signals
training
network model
sound signal
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2018/095377
Other languages
English (en)
French (fr)
Inventor
丁惠君
潘搌鹏
岑乾
但果
汪天富
陈思平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen University
Original Assignee
Shenzhen University
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 Shenzhen University filed Critical Shenzhen University
Publication of WO2020000523A1 publication Critical patent/WO2020000523A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
    • G10L25/66—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination for extracting parameters related to health condition
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination

Definitions

  • the present invention relates to the field of signal processing, and in particular, to a signal processing method and device.
  • the human voice When the volume, pitch, sound quality, duration of pronunciation, resonance position, and ease of pronunciation of the human voice are abnormal, the human voice may not meet the needs of people's daily life and work. At this time, the human voice is considered to have a voice disease. . Since the voice plays a very important role in the process of language formation and use, how to determine the voice disease has become an urgent technical problem. At present, people's voice diseases can be checked by dynamic laryngoscopes, laryngeal electromyography, etc., but such technical methods are invasive examination methods, which have certain pain and injury for patients.
  • Embodiments of the present invention provide a signal processing method and device, which can detect voice diseases through non-contact.
  • a first aspect of the embodiments of the present invention provides a signal processing method, including:
  • the method further includes:
  • An initial network model is trained using the training sound signal to obtain a target network model.
  • the method further includes:
  • the training an initial network model using the training sound signal to obtain a target network model includes:
  • An initial network model is trained using the divided training signals to obtain a target network model.
  • the method further includes:
  • the probability that the detected sound signal belongs to each type of signal through the target network model includes:
  • the segmented signal is input to a target network model to obtain a probability that the detected sound signal belongs to each type of signal.
  • the method further includes:
  • Dividing each sound signal in the training sound signal into signals of the same length, and obtaining the divided training signal includes:
  • Each signal in the partial signals is divided into signals of the same length to obtain a divided training signal.
  • a second aspect of the embodiments of the present invention provides a signal processing apparatus, including:
  • a first acquisition unit configured to acquire and detect a sound signal through a microphone
  • a recognition unit configured to recognize, by using a target network model, a probability that a detection sound signal collected by the first acquisition unit belongs to each type of signal;
  • the classification unit is configured to classify the detection sound signal collected by the first acquisition unit to a signal class with the highest probability identified by the identification unit.
  • the apparatus further includes:
  • a second acquisition unit configured to acquire a training sound signal through a microphone
  • a training unit is configured to train an initial network model using a training sound signal collected by the second acquisition unit to obtain a target network model.
  • the apparatus further includes:
  • a first division unit configured to divide each sound signal in the training sound signal collected by the second acquisition unit into signals of the same length to obtain a divided training signal
  • the training unit is specifically configured to train an initial network model using the divided training signals obtained by the first dividing unit to obtain a target network model.
  • the apparatus further includes:
  • a second division unit configured to divide the detection sound signal collected by the first acquisition unit into signals of the same length to obtain segmented signals, each signal in the segmented signal and each signal in the divided training signal Are the same length;
  • the identification unit is specifically configured to input a segmented signal obtained by the second division unit into a target network model, and obtain a probability that the detected sound signal belongs to each type of signal.
  • the apparatus further includes:
  • a selection unit configured to select a partial signal from each of the training sound signals collected by the second acquisition unit
  • the first dividing unit is specifically configured to divide each of the signals selected by the selecting unit into signals of the same length to obtain a divided training signal.
  • a third aspect of the embodiments of the present invention provides a signal processing device.
  • the signal processing device includes a processor, a memory, and a microphone.
  • the memory is used to store program code
  • the processor is used to execute the program code
  • the microphone is used to collect sound signals.
  • the processor executes the program code stored in the memory, the processor is caused to execute the signal processing method disclosed in the first aspect or any possible implementation manner of the first aspect.
  • a fourth aspect of the embodiments of the present invention provides a readable storage medium that stores a program for a signal processing apparatus to execute the signal processing method disclosed in the first aspect or any possible implementation manner of the first aspect. Code.
  • a detection sound signal is collected through a microphone, and a probability that the detection sound signal belongs to each type of signal is recognized through a target network model, and the detection sound signal is classified into a signal type with a highest probability. Because the human voice signal is collected through the microphone, it can be kept out of contact with people, so that it does not cause pain and injury to the patient, so that voice diseases can be detected through non-contact.
  • FIG. 1 is a schematic flowchart of a signal processing method according to an embodiment of the present invention
  • FIG. 2 is a schematic flowchart of another signal processing method according to an embodiment of the present invention.
  • FIG. 3 is a schematic structural diagram of a signal processing device according to an embodiment of the present invention.
  • FIG. 4 is a schematic structural diagram of another signal processing apparatus according to an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of a residual CNN model according to an embodiment of the present invention.
  • Embodiments of the present invention provide a signal processing method and device, which can detect voice diseases through non-contact. Each of them will be described in detail below.
  • FIG. 1 is a schematic flowchart of a signal processing method according to an embodiment of the present invention. As shown in FIG. 1, the signal processing method may include the following steps.
  • the sound signals can be divided into two types, one is the sound signal of a normal person, and the other is the sound signal of a voice patient with a diseased throat. Sound signals can also be classified into three or more categories, and voice patients with throat diseases can be divided into multiple categories according to the severity of the disease.
  • the collection frequency of the microphone may be 8kHZ, or other values, which is not limited in this embodiment. In order to ensure the accuracy of the collected sound signals, the distance between the microphone and the human lips should be less than 30cm.
  • the collected detection sound signal may include only one tone signal, or may include multiple tone signals.
  • the collected detection sound signal may be one sound signal or multiple sound signals. When there are multiple sound signals, the lengths of the multiple sound signals may be the same or different.
  • the target network model is a pre-trained network model, which can include a Convolutional Neural Network (CNN) model and a Recurrent Neural Network (RNN) model, which can be a pre-CNN and a RNN.
  • the CNN model may be a residual CNN model, or an inception module in a Google network model, or a dense module in a dense network model.
  • the RNN model can be a long short-term memory (LSTM) RNN model, or a gated recurrent unit (GRU) model.
  • LSTM long short-term memory
  • GRU gated recurrent unit
  • FIG. 5 is a schematic diagram of a residual CNN model according to an embodiment of the present invention.
  • the input of the residual CNN model is outputted after operations such as a convolution layer, a batch normalization layer, a ReLU function, and a superposition.
  • the target network model may also include a normalized probability (softmax) function, a maxpooling algorithm, and the like.
  • the detected sound signal is classified into the signal category with the highest probability.
  • the category with the highest probability is a normal person, it indicates that the detected sound signal is Normal human voice signal; when the category with the highest probability is the voice signal of the voice patient, it indicates that the detected voice signal is the voice signal of the voice patient.
  • a detection sound signal is collected through a microphone, a probability that the detection sound signal belongs to each type of signal is recognized through a target network model, and the detection sound signal is classified into a signal type with a highest probability. Because the human voice signal is collected through the microphone, it can be kept out of contact with people, so that it does not cause pain and injury to the patient, so that voice diseases can be detected through non-contact.
  • FIG. 2 is a schematic flowchart of another signal processing method according to an embodiment of the present invention. As shown in FIG. 2, the signal processing method may include the following steps.
  • a training sound signal in order to detect a human voice disease, a training sound signal needs to be collected through a microphone first.
  • the training sound signal includes multiple sound signals, and the lengths of the multiple sound signals may be all the same, or they may all be different, or may be partially the same or different.
  • the distance between the microphone and the human lips should be less than 30cm.
  • the ambient noise in order to reduce the impact of ambient noise on sound signal detection, the ambient noise should be lower than 30db.
  • Each of the training sound signals may include only one tone signal or multiple tone signals.
  • the training sound signal may be first extracted from the training sound signal. Select a part of each of the sound signals. This part of the signal can be a 50% -70% signal randomly selected from the original sound signal.
  • a part of the signals is selected from each of the training sound signals for the first round of training, and then the training sound signals are selected from the training sound signals. For each of the sound signals, select another part of the signal for the second round of training until the end of the training.
  • the length of a part of the signals selected from the same sound signal in each round of selection training is the same, but the content may be different or the same part may be different.
  • each signal in the partial signal is divided into signals of the same length, that is, the partial signal is subjected to frame processing to obtain division training. signal.
  • the length of each frame is the same.
  • the number of frames obtained after framing each part of the signal is different.
  • each signal in the training sound signal can be directly divided into signals of the same length to obtain divided training. Signal without performing step 202.
  • the formula for dividing training signals can be expressed as follows:
  • y (n) is the divided training signal, that is, the set of divided signals
  • h (m) is the Hamming window function
  • x (n) is a part of the training signal with the same length as the window function.
  • N is the length of each segment of the signal and the Hamming window function (or window) after the segmentation
  • m is the point inside the window
  • M is the overlap rate, generally 0.5
  • n is a natural number.
  • n 0, ..., P-1
  • P is the number of signals in the divided training signal.
  • each of the partial signals is divided into signals of the same length or each of the training sound signals is divided into signals of the same length.
  • the initial network model is trained by using the divided training signals. Obtain the target network model.
  • the signals belonging to the first signal in the divided training signals can be input to the initial network model to obtain the probability that the first signal belongs to each type of signal, a loss function value is calculated according to the probability, and the loss function value is propagated back to the initial network model to obtain Update the initial network model of the parameters, and then divide the signal belonging to the second signal in the training signal into the initial network model of the update parameters to obtain the probability that the second signal belongs to each type of signal.
  • the values are propagated back to the initial network model with the updated parameters, to obtain the initial network model with the updated parameters again, and the above steps are repeated until the minimum loss function value is determined, and the initial network model corresponding to the minimum loss function value is used to determine the target network. model.
  • the first signal and the second signal are different signals in the training sound signal.
  • the initial network model is the same as the target network model, but the parameters are different.
  • the initial network model and the target network model may include a CNN model and an RNN model, and may be a front CNN and a RNN.
  • the CNN model can be a residual CNN model, an initial module in a Google network model, or a dense module in a dense network model.
  • the RNN model can be an LSTMRNN model or a GRU model.
  • the initial network model and the target network model may also include a normalized probability function, a maximum pooling algorithm, and the like.
  • the initial network model and the target network model can be understood as the input vector x multiplied by the network weight w, and then obtained by a non-linear transformation f and adding a bias b.
  • s can be expressed as follows:
  • p can be expressed as follows:
  • j 0 or 1
  • k 0,1.
  • the loss function can be expressed as follows:
  • loss (p, label) represents the loss function
  • label represents the true label of the data.
  • Step 205 is the same as step 101.
  • Step 101 For detailed description, refer to step 101, and details are not described herein again.
  • Step 206 is similar to step 203.
  • Step 203 For a detailed description, refer to step 203, and details are not described herein again.
  • Step 208 is similar to step 103.
  • Step 103 For a detailed description, refer to step 103, and details are not described herein again.
  • FIG. 3 is a schematic structural diagram of a signal processing device according to an embodiment of the present invention.
  • the signal processing device can be applied to the signal processing method shown in FIG. 1 to FIG. 2.
  • the device may include:
  • a first acquisition unit 301 configured to acquire and detect a sound signal through a microphone
  • An identification unit 302 configured to identify, by using a target network model, a probability that a detection sound signal collected by the first acquisition unit 301 belongs to each type of signal;
  • the classification unit 303 is configured to classify the detection sound signal collected by the first acquisition unit 301 to a signal class with the highest probability identified by the identification unit 302.
  • the apparatus may further include:
  • a second acquisition unit 304 configured to acquire a training sound signal through a microphone
  • the training unit 305 is configured to train an initial network model using a training sound signal collected by the second acquisition unit 304 to obtain a target network model.
  • the recognition unit 302 is configured to recognize, through the target network model obtained by the training unit 305, a probability that a sound signal belongs to each type of signal.
  • the apparatus may further include:
  • a first division unit 306, configured to divide each sound signal in the training sound signal collected by the second acquisition unit 304 into signals of the same length to obtain a divided training signal
  • the training unit 305 is specifically configured to train the initial network model using the divided training signals obtained by the first dividing unit 306 to obtain a target network model.
  • the apparatus may further include:
  • a second dividing unit 307 configured to divide the sound signal collected by the first collecting unit 301 into signals of the same length to obtain segmented signals, where each signal in the segmented signal has the same length as each signal in the divided training signal;
  • the identifying unit 302 is specifically configured to input the segmented signal obtained by the second dividing unit 307 into a target network model, and obtain a probability that the detected sound signal belongs to each type of signal.
  • the apparatus may further include:
  • a selecting unit 308, configured to select a partial signal from each of the training sound signals collected by the second collecting unit 304;
  • the first dividing unit 306 is specifically configured to divide each of the signals selected by the selecting unit 308 into signals of the same length to obtain a divided training signal.
  • FIG. 4 is a schematic structural diagram of another signal processing apparatus according to an embodiment of the present invention.
  • the signal processing device can be applied to the signal processing method shown in FIG. 1 to FIG. 2.
  • the signal processing device may include: at least one processor 401, such as a CPU, a memory 402, a microphone 403, a graphics processing unit (Graphics Processing Unit, GPU) 404, and at least one bus 405.
  • the memory 402 may be a high-speed RAM memory, or may be a non-volatile memory (non-volatile memory), for example, at least one disk memory.
  • the memory 402 may also be at least one storage device located far from the foregoing processor 401. among them:
  • the bus 405 is configured to implement connection and communication between these components;
  • a microphone 403, configured to collect a detection sound signal and send it to the processor 401;
  • a set of program code is stored in the memory 402, and the processor 401 is configured to call the program code stored in the memory 402 to perform the following operations:
  • the detection sound signal is classified into a signal class with the highest probability.
  • the microphone 403 is further configured to collect a training sound signal through the microphone;
  • the processor 401 is further configured to call the program code stored in the memory 402 to perform the following operations:
  • the training network is used to train the initial network model to obtain the target network model.
  • the processor 401 is further configured to call the program code stored in the memory 402 to perform the following operations:
  • the processor 401 uses the training sound signal to train the initial network model.
  • the target network model obtained includes:
  • the initial network model is trained using the divided training signals to obtain the target network model.
  • the processor 401 is further configured to call the program code stored in the memory 402 to perform the following operations:
  • the probability that the processor 401 recognizes that the sound signal belongs to each type of signal through the target network model includes:
  • the segmented signal is input into the target network model to obtain the probability that the detected sound signal belongs to each type of signal.
  • the processor 401 is further configured to call the program code stored in the memory 402 to perform the following operations:
  • the processor 401 divides each sound signal in the training sound signal into signals of the same length, and obtaining the divided training signal includes:
  • Each signal in this partial signal is divided into signals of the same length to obtain a divided training signal.
  • steps 101, 201, and 205 can be performed by the microphone 403 in the signal processing device, and steps 102-step 103, step 202-step 204, and step 206-step 208 can be performed by the processor 401 and Memory 402 to execute.
  • the acquisition unit 301 may be implemented by an acceleration sensor 403 in the signal processing device, and the recognition unit 302, the classification unit 303, the training unit 304, the first division unit 305, the second division unit 306, and the selection unit 307 may be processed by a signal. It is implemented by a processor 401 and a memory 402 in the device.
  • An embodiment of the present invention also discloses a readable storage medium that stores a program code of a signal processing apparatus for executing the signal processing method shown in FIGS. 1-2.
  • the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium.
  • the technical solution of the present invention is essentially a part that contributes to the existing technology, or all or part of the technical solution may be embodied in the form of a software product, which is stored in a storage medium. Included are several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in the embodiments of the present invention.
  • the foregoing storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks and other media that can store program codes .

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Public Health (AREA)
  • General Health & Medical Sciences (AREA)
  • Epidemiology (AREA)
  • Circuit For Audible Band Transducer (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

本发明实施例提供了一种信号处理方法及装置，包括：通过麦克风采集检测声音信号；通过目标网络模型识别该检测声音信号属于每类信号的概率；将该检测声音信号归类至识别出的概率最大的信号类。本发明实施例，可以通过非接触式检查嗓音疾病。

Description

一种信号处理方法及装置 技术领域
本发明涉及信号处理领域，具体涉及一种信号处理方法及装置。
背景技术
当人嗓音的音量、音调、音质、发音持续时间、共鸣位置以及发音的轻松程度等出现异常时，人的声音可能无法满足人们日常生活和工作的需要，此时认为人的嗓子患有嗓音疾病。由于在语言形成及运用过程中，嗓音起着非常重要的作用，因此，如何确定嗓音疾病已成为一个亟待解决的技术问题。目前，可以通过动态喉镜、喉肌电图等检查人们的嗓音疾病，但此类技术手段均属于侵入性检查方式，对于患者存在一定的痛苦及损伤。
发明内容
本发明实施例提供一种信号处理方法及装置，可以通过非接触式检测嗓音疾病。
本发明实施例第一方面提供一种信号处理方法，包括：
通过麦克风采集检测声音信号；
通过目标网络模型识别所述检测声音信号属于每类信号的概率；
将所述检测声音信号归类至所述概率最大的信号类。
在一个实施例中，所述方法还包括：
通过麦克风采集训练声音信号；
使用所述训练声音信号训练初始网络模型，获得目标网络模型。
在一个实施例中，所述方法还包括：
将所述训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号；
所述使用所述训练声音信号训练初始网络模型，获得目标网络模型包括：
使用所述划分训练信号训练初始网络模型，获得目标网络模型。
在一个实施例中，所述方法还包括：
将所述检测声音信号划分为长度相同的信号，获得分段信号，所述分段信号中每个信号与所述划分训练信号中每个信号的长度相同；
所述通过目标网络模型识别所述检测声音信号属于每类信号的概率包括：
将所述分段信号输入目标网络模型，获得所述检测声音信号属于每类信号的概率。
在一个实施例中，所述方法还包括：
从所述训练声音信号中的每个声音信号中选取部分信号；
所述将所述训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号包括：
将所述部分信号中每个信号划分为长度相同的信号，获得划分训练信号。
本发明实施例第二方面提供一种信号处理装置，包括：
第一采集单元，用于通过麦克风采集检测声音信号；
识别单元，用于通过目标网络模型识别所述第一采集单元采集的检测声音信号属于每类信号的概率；
归类单元，用于将所述第一采集单元采集的检测声音信号归类至所述识别单元识别出的概率最大的信号类。
在一个实施例中，所述装置还包括：
第二采集单元，用于通过麦克风采集训练声音信号；
训练单元，用于使用所述第二采集单元采集的训练声音信号训练初始网络模型，获得目标网络模型。
在一个实施例中，所述装置还包括：
第一划分单元，用于将所述第二采集单元采集的训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号；
所述训练单元，具体用于使用所述第一划分单元获得的划分训练信号训练初始网络模型，获得目标网络模型。
在一个实施例中，所述装置还包括：
第二划分单元，用于将所述第一采集单元采集的检测声音信号划分为长度 相同的信号，获得分段信号，所述分段信号中每个信号与所述划分训练信号中每个信号的长度相同；
所述识别单元，具体用于将所述第二划分单元获得的分段信号输入目标网络模型，获得所述检测声音信号属于每类信号的概率。
在一个实施例中，所述装置还包括：
选取单元，用于从所述第二采集单元采集的训练声音信号中的每个声音信号中选取部分信号；
所述第一划分单元，具体用于将所述选取单元选取的部分信号中每个信号划分为长度相同的信号，获得划分训练信号。
本发明实施例第三方面提供一种信号处理装置，该信号处理装置包括处理器、存储器和麦克风，存储器用于存储程序代码，处理器用于执行程序代码，麦克风用于采集声音信号。当处理器执行存储器存储的程序代码时，使得处理器执行第一方面或第一方面的任一种可能实现方式所公开的信号处理方法。
本发明实施例第四方面提供一种可读存储介质，该可读存储介质存储了信号处理装置用于执行第一方面或第一方面的任一种可能实现方式所公开的信号处理方法的程序代码。
本发明实施例中，通过麦克风采集检测声音信号，通过目标网络模型识别该检测声音信号属于每类信号的概率，将该检测声音信号归类至概率最大的信号类。由于通过麦克风采集人声音信号，因此，可以不与人接触，从而不会对患者存造成痛苦及损伤，以便通过非接触式检测嗓音疾病。
附图说明
为了更清楚地说明本发明实施例中的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本发明实施例提供的一种信号处理方法的流程示意图；
图2是本发明实施例提供的另一种信号处理方法的流程示意图；
图3是本发明实施例提供的一种信号处理装置的结构示意图；
图4是本发明实施例提供的另一种信号处理装置的结构示意图；
图5是本发明实施例提供的一种残差CNN模型的示意图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
本发明实施例提供一种信号处理方法及装置，可以通过非接触式检测嗓音疾病。以下分别进行详细说明。
请参阅图1，图1是本发明实施例提供的一种信号处理方法的流程示意图。如图1所示，该信号处理方法可以包括以下步骤。
101、通过麦克风采集检测声音信号。
本实施例中，当需要检查人的嗓子是否患有疾病时，先需要通过麦克风采集检测声音信号。声音信号可以分为两类，一类为正常人的声音信号，另一类为嗓子患有疾病的嗓音患者的声音信号。声音信号也可以分为三类或三类以上，可以将嗓子患有疾病的嗓音患者根据患病情况的严重程度不同分为多类。麦克风的采集频率可以为8kHZ，也可以为其它值，本实施例不作限定。为了保证采集的声音信号的准确性，麦克风与人的嘴唇之间的距离要小于30cm。此外，为了减少环境噪音对声音信号检测的影响，环境噪音要低于30db。采集的检测声音信号可以只包括一个音的信号，也可以包括多个音的信号。此外，采集的检测声音信号可以是一个声音信号，也可以是多个声音信号。当为多个声音信号时，这多个声音信号的长度可以相同，也可以不同。
102、通过目标网络模型识别检测声音信号属于每类信号的概率。
本实施例中，为了让采集到的声音信号尽可能地保留高频特征，可以不对采集的声音信号进行滤波处理。通过麦克风采集到需要的检测声音信号之后，通过目标网络模型识别检测声音信号属于每类信号的概率。目标网络模型是预 先训练好的网络模型，可以包括卷积神经网络(Convolutional Neural Network，CNN)模型和循环神经网络(Recurrent Neural Network，RNN)模型，可以为前CNN后RNN。CNN模型可以为残差CNN模型，也可以为google网络模型中的初始模块(inception module)，还可以为密集网络(densenet)模型中的密集模块(dense module)。RNN模型可以为长短期记忆(Long Short Time Memory，LSTM)RNN模型，也可以为门控循环单元(Gated Recurrent Unit，GRU)模型。请参阅图5，图5是本发明实施例提供的一种残差CNN模型的示意图。如图5所示，残差CNN模型的输入经过卷积层、批量归一化层、ReLU函数、叠加等运算之后输出。此外，目标网络模型还可以包括归一化概率(softmax)函数、最大池化(maxpooling)算法等。
103、将检测声音信号归类至概率最大的信号类。
本实施例中，通过目标网络模型识别出检测声音信号属于每类信号的概率之后，将检测声音信号归类至概率最大的信号类，当概率最大的类别为正常人时，表明检测声音信号为正常人的声音信号；当概率最大的类别为嗓音患者的声音信号时，表明检测声音信号为嗓音患者的声音信号。
在图1所描述的信号处理方法中，通过麦克风采集检测声音信号，通过目标网络模型识别检测声音信号属于每类信号的概率，将检测声音信号归类至概率最大的信号类。由于通过麦克风采集人声音信号，因此，可以不与人接触，从而不会对患者存造成痛苦及损伤，以便通过非接触式检测嗓音疾病。
请参阅图2，图2是本发明实施例提供的另一种信号处理方法的流程示意图。如图2所示，该信号处理方法可以包括以下步骤。
201、通过麦克风采集训练声音信号。
本实施例中，为了检测人嗓音疾病，需要先通过麦克风采集训练声音信号。训练声音信号包括多个声音信号，这多个声音信号的长度可以均相同，也可以均不同，还可以部分相同部分不同。为了保证采集的声音信号的准确性，麦克风与人的嘴唇之间的距离要小于30cm。此外，为了减少环境噪音对声音信号检测的影响，环境噪音要低于30db。训练声音信号中的每个声音信号可以只包括一个音的信号，也可以包括多个音的信号。
202、从训练声音信号中的每个声音信号中选取部分信号。
本实施例中，通过麦克风采集到训练声音信号之后，当采集的训练声音信号中声音信号的数量较少时，为了减轻过拟合，增加训练的声音信号的数量，可以先从训练声音信号中的每个声音信号中选取部分信号。这部分信号可以是从原声音信号中随机选取的50％-70％的信号。
本实施例中，当采集的训练声音信号中声音信号的数量较少时，一般是先从训练声音信号中的每个声音信号中分别选取一部分信号进行第一轮训练，之后再从训练声音信号中的每个声音信号中分别选取另一部分信号进行第二轮训练，直到训练结束。其中，每轮选训练中从同一声音信号中选取的部分信号的长度均相同，但内容可以均不同，也可以部分相同部分不同。
203、将该部分信号中每个信号划分为长度相同的信号，获得划分训练信号。
本实施例中，从训练声音信号中的每个声音信号中选取出部分信号之后，将该部分信号中每个信号划分为长度相同的信号，即对该部分信号进行分帧处理，获得划分训练信号。其中，每一帧的长度相同，当部分信号的长度不同时，每部分信号分帧之后得到的帧数不同。在一个实施例中，当采集的训练声音信号中声音信号的数量较大时，通过麦克风采集到训练声音信号之后，可以直接将训练声音信号中每个信号划分为长度相同的信号，获得划分训练信号，而不需要执行步骤202。其中，划分训练信号的公式可以表示如下：
y(n)＝∏x(n)h(m)
其中，y(n)为划分训练信号，即划分后信号的集合，h(m)为汉明窗函数，x(n)为长度与窗函数的长度相同的训练信号中的部分信号。h(m)可以表示如下：
Figure PCTCN2018095377-appb-000001
其中，N为分段后每段信号和汉明窗函数(或窗)的长度，m为窗里面的点，0≤m≤N-1，M为重叠率，一般为0.5，n为自然数，n＝0，……，P-1，P为划分训练信号中信号的数量。
204、使用划分训练信号训练初始网络模型，获得目标网络模型。
本实施例中，将该部分信号中每个信号划分为长度相同的信号或将训练声音信号中每个信号划分为长度相同的信号，获得划分训练信号之后，使用划分训练信号训练初始网络模型，获得目标网络模型。可以将划分训练信号中属于第一信号的信号输入初始网络模型，获得第一信号属于每类信号的概率，根据该概率计算损失函数值，将该损失函数值反向传播至初始网络模型，获得更新参数的初始网络模型，之后将划分训练信号中属于第二信号的信号输入更新参数的初始网络模型，获得第二信号属于每类信号的概率，根据该概率计算损失函数值，将该损失函数值反向传播至更新参数的初始网络模型，获得再次更新参数的初始网络模型，重复执行上述步骤，直到确定出最小的损失函数值，并将损失函数值最小时对应的初始网络模型确定目标网络模型。第一信号和第二信号为训练声音信号中的不同信号。其中，初始网络模型与目标网络模型的模型相同，只是参数不同。初始网络模型和目标网络模型可以包括CNN模型和RNN模型，可以为前CNN后RNN。CNN模型可以为残差CNN模型，也可以为google网络模型中的初始模块，还可以为密集网络模型中的密集模块。RNN模型可以为LSTMRNN模型，也可以为GRU模型。此外，初始网络模型和目标网络模型还可以包括归一化概率函数、最大池化算法等。
本实施例中，当声音信号分为两类时，初始网络模型和目标网络模型可以理解为输入向量x与网络权重w相乘，之后经过一个非线性变换f，并加上偏置b后得到的二维向量s，再对s进行归一化(即经过softmax函数)后得到的两个类别的概率p。其中，s可以表示如下：
s＝f(wx)+b
p可以表示如下：
Figure PCTCN2018095377-appb-000002
其中，j为声音信号的类别，当声音信号分为两类时，j＝0或1，k＝0，1。损失函数可以表示如下：
Figure PCTCN2018095377-appb-000003
loss(p，label)表示损失函数，label表示数据的真实标签。将损失函数值反向传播至初始网络模型获得更新参数的初始网络模型，可以为损失函数对每个变量求偏导，之后计算所有偏导的平均值，使用初始网络模型里面的参数均减去该平均值，获得更新参数的初始网络模型。
205、通过麦克风采集检测声音信号。
其中，步骤205与步骤101相同，详细描述请参考步骤101，在此不再赘述。
206、将检测声音信号划分为长度相同的信号，获得分段信号。
其中，步骤206与步骤203相似，详细描述请参考步骤203，在此不再赘述。
207、将分段信号输入目标网络模型，获得检测声音信号属于每类信号的概率。
208、将检测声音信号归类至概率最大的信号类。
其中，步骤208与步骤103相似，详细描述请参考步骤103，在此不再赘述。
在图2所描述的信号处理方法中，由于通过麦克风采集人声音信号，因此，可以不与人接触，从而不会对患者存造成痛苦及损伤，以便通过非接触式检测嗓音疾病。
基于上述实施例中的信号处理方法的同一构思，请参阅图3，图3是本发明实施例提供的一种信号处理装置的结构示意图。其中，该信号处理装置可以应用于上述图1-图2所示的信号处理方法中。如图3所示，该装置可以包括：
第一采集单元301，用于通过麦克风采集检测声音信号；
识别单元302，用于通过目标网络模型识别第一采集单元301采集的检测声音信号属于每类信号的概率；
归类单元303，用于将第一采集单元301采集的检测声音信号归类至识别单元302识别出的概率最大的信号类。
作为一种可能的实施方式，该装置还可以包括：
第二采集单元304，用于通过麦克风采集训练声音信号；
训练单元305，用于使用第二采集单元304采集的训练声音信号训练初始网络模型，获得目标网络模型。
具体地，识别单元302，用于通过训练单元305获得的目标网络模型识别声 音信号属于每类信号的概率。
作为一种可能的实施方式，该装置还可以包括：
第一划分单元306，用于将第二采集单元304采集的训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号；
训练单元305，具体用于使用第一划分单元306获得的划分训练信号训练初始网络模型，获得目标网络模型。
作为一种可能的实施方式，该装置还可以包括：
第二划分单元307，用于将第一采集单元301采集的声音信号划分为长度相同的信号，获得分段信号，分段信号中每个信号与划分训练信号中每个信号的长度相同；
识别单元302，具体用于将第二划分单元307获得的分段信号输入目标网络模型，获得检测声音信号属于每类信号的概率。
作为一种可能的实施方式，该装置还可以包括：
选取单元308，用于从第二采集单元304采集的训练声音信号中的每个声音信号中选取部分信号；
第一划分单元306，具体用于将选取单元308选取的部分信号中每个信号划分为长度相同的信号，获得划分训练信号。
有关上述采集单元301、识别单元302、归类单元303、训练单元304、第一划分单元305、第二划分单元306和选取单元307更详细的描述可以直接参考上述图1-图2所示的信号处理方法实施例中的相关描述直接得到，这里不加赘述。
基于上述实施例中的信号处理方法的同一构思，请参阅图4，图4是本发明实施例提供的另一种信号处理装置的结构示意图。其中，该信号处理装置可以应用于上述图1-图2所示的信号处理方法中。如图4所示，该信号处理装置可以包括：至少一个处理器401，如CPU，存储器402，麦克风403、图像处理器(Graphics Processing Unit，GPU)404以及至少一个总线405。存储器402可以是高速RAM存储器，也可以是非不稳定的存储器(non-volatile memory)，例如至少一个磁盘存储器。可选地，存储器402还可以是至少一个位于远离前述处理器401的存储装置。其中：
总线405，用于实现这些组件之间的连接通信；
GPU404，用于辅助浮点计算；
麦克风403，用于采集检测声音信号，并发送给处理器401；
存储器402中存储有一组程序代码，处理器401用于调用存储器402中存储的程序代码执行以下操作：
通过目标网络模型识别检测声音信号属于每类信号的概率；
将检测声音信号归类至概率最大的信号类。
作为一种可能的实施方式，麦克风403，还用于通过麦克风采集训练声音信号；
处理器401还用于调用存储器402中存储的程序代码执行以下操作：
使用训练声音信号训练初始网络模型，获得目标网络模型。
作为一种可能的实施方式，处理器401还用于调用存储器402中存储的程序代码执行以下操作：
将训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号；
处理器401使用训练声音信号训练初始网络模型，获得目标网络模型包括：
使用划分训练信号训练初始网络模型，获得目标网络模型。
作为一种可能的实施方式，处理器401还用于调用存储器402中存储的程序代码执行以下操作：
将检测声音信号划分为长度相同的信号，获得分段信号，分段信号中每个信号与划分训练信号中每个信号的长度相同；
处理器401通过目标网络模型识别检测声音信号属于每类信号的概率包括：
将分段信号输入目标网络模型，获得检测声音信号属于每类信号的概率。
作为一种可能的实施方式，处理器401还用于调用存储器402中存储的程序代码执行以下操作：
从训练声音信号中的每个声音信号中选取部分信号；
处理器401将训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号包括：
将该部分信号中每个信号划分为长度相同的信号，获得划分训练信号。
其中，步骤101、步骤201和步骤205可以由信号处理装置中的麦克风403来执行，步骤102-步骤103、步骤202-步骤204以及步骤206-步骤208可以由信号处理装置中的处理器401和存储器402来执行。
其中，采集单元301可以由信号处理装置中的加速度传感器403来实现，识别单元302、归类单元303、训练单元304、第一划分单元305、第二划分单元306和选取单元307可以由信号处理装置中的处理器401和存储器402来实现。
本发明实施例还公开了一种可读存储介质，该可读存储介质存储了信号处理装置用于执行图1-图2所示的信号处理方法的程序代码。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分，或者该技术方案的全部或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一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本发明各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
以上对本发明实施例提供的信号处理方法及装置进行了详细介绍，本文中应用了具体个例对本发明的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本发明的方法及其核心思想；同时，对于本领域的一般技术人员，依据本发明的思想，在具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本发明的限制。

Claims (16)

  1. 一种信号处理方法，其特征在于，包括：
    通过麦克风采集检测声音信号；
    通过目标网络模型识别所述检测声音信号属于每类信号的概率；
    将所述检测声音信号归类至所述概率最大的信号类。
  2. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    通过麦克风采集训练声音信号；
    使用所述训练声音信号训练初始网络模型，获得目标网络模型。
  3. 根据权利要求2所述的方法，其特征在于，所述方法还包括：
    将所述训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号；
    所述使用所述训练声音信号训练初始网络模型，获得目标网络模型包括：
    使用所述划分训练信号训练初始网络模型，获得目标网络模型。
  4. 根据权利要求3所述的方法，其特征在于，所述方法还包括：
    将所述检测声音信号划分为长度相同的信号，获得分段信号，所述分段信号中每个信号与所述划分训练信号中每个信号的长度相同；
    所述通过目标网络模型识别所述检测声音信号属于每类信号的概率包括：
    将所述分段信号输入目标网络模型，获得所述检测声音信号属于每类信号的概率。
  5. 根据权利要求3所述的方法，其特征在于，所述方法还包括：
    从所述训练声音信号中的每个声音信号中选取部分信号；
    所述将所述训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号包括：
    将所述部分信号中每个信号划分为长度相同的信号，获得划分训练信号。
  6. 一种信号处理装置，其特征在于，包括：
    第一采集单元，用于通过麦克风采集检测声音信号；
    识别单元，用于通过目标网络模型识别所述第一采集单元采集的检测声音信号属于每类信号的概率；
    归类单元，用于将所述第一采集单元采集的检测声音信号归类至所述识别单元识别出的概率最大的信号类。
  7. 根据权利要求6所述的装置，其特征在于，所述装置还包括：
    第二采集单元，用于通过麦克风采集训练声音信号；
    训练单元，用于使用所述第二采集单元采集的训练声音信号训练初始网络模型，获得目标网络模型。
  8. 根据权利要求7所述的方法，其特征在于，所述装置还包括：
    第一划分单元，用于将所述第二采集单元采集的训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号；
    所述训练单元，具体用于使用所述第一划分单元获得的划分训练信号训练初始网络模型，获得目标网络模型。
  9. 根据权利要求8所述的装置，其特征在于，所述装置还包括：
    第二划分单元，用于将所述第一采集单元采集的检测声音信号划分为长度相同的信号，获得分段信号，所述分段信号中每个信号与所述划分训练信号中每个信号的长度相同；
    所述识别单元，具体用于将所述第二划分单元获得的分段信号输入目标网络模型，获得所述检测声音信号属于每类信号的概率。
  10. 根据权利要求8所述的装置，其特征在于，所述装置还包括：
    选取单元，用于从所述第二采集单元采集的训练声音信号中的每个声音信号中选取部分信号；
    所述第一划分单元，具体用于将所述选取单元选取的部分信号中每个信号 划分为长度相同的信号，获得划分训练信号。
  11. 一种信号处理装置，其特征在于，包括处理器、存储器和麦克风，其中：
    所述麦克风，用于采集检测声音信号；
    所述存储器中存储有一组程序代码，所述处理器用于调用所述存储器中存储的程序代码执行以下操作：
    通过目标网络模型识别所述检测声音信号属于每类信号的概率；
    将所述检测声音信号归类至所述概率最大的信号类。
  12. 根据权利要求11所述的装置，其特征在于，所述麦克风，还用于采集训练声音信号；
    所述处理器还用于调用所述存储器中存储的程序代码执行以下操作：
    使用所述训练声音信号训练初始网络模型，获得目标网络模型。
  13. 根据权利要求12所述的装置，其特征在于，所述处理器还用于调用所述存储器中存储的程序代码执行以下操作：
    将所述训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号；
    所述处理器使用所述训练声音信号训练初始网络模型，获得目标网络模型包括：
    使用所述划分训练信号训练初始网络模型，获得目标网络模型。
  14. 根据权利要求13所述的装置，其特征在于，所述处理器还用于调用所述存储器中存储的程序代码执行以下操作：
    将所述检测声音信号划分为长度相同的信号，获得分段信号，所述分段信号中每个信号与所述划分训练信号中每个信号的长度相同；
    所述处理器通过目标网络模型识别所述检测声音信号属于每类信号的概率包括：
    将所述分段信号输入目标网络模型，获得所述检测声音信号属于每类信号的概率。
  15. 根据权利要求13所述的装置，其特征在于，所述处理器还用于调用所述存储器中存储的程序代码执行以下操作：
    从所述训练声音信号中的每个声音信号中选取部分信号；
    所述处理器将所述训练声音信号中每个声音信号划分为长度相同的信号，获得划分训练信号包括：
    将所述部分信号中每个信号划分为长度相同的信号，获得划分训练信号。
  16. 一种计算机可读存储介质，其特征在于，所述存储介质上存储有程序，当程序运行时，实现如权利要求1-5任一项所述的信号处理方法。
PCT/CN2018/095377 2018-06-26 2018-07-12 一种信号处理方法及装置 Ceased WO2020000523A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810669024.4 2018-06-26
CN201810669024.4A CN109192226A (zh) 2018-06-26 2018-06-26 一种信号处理方法及装置

Publications (1)

Publication Number Publication Date
WO2020000523A1 true WO2020000523A1 (zh) 2020-01-02

Family

ID=64948488

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/095377 Ceased WO2020000523A1 (zh) 2018-06-26 2018-07-12 一种信号处理方法及装置

Country Status (2)

Country Link
CN (1) CN109192226A (zh)
WO (1) WO2020000523A1 (zh)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109758141B (zh) * 2019-03-06 2022-04-26 清华大学 一种心理压力监测方法、装置及系统
CN111179972A (zh) * 2019-12-12 2020-05-19 中山大学 一种基于深度学习的人声检测算法
CN114373452A (zh) * 2020-10-15 2022-04-19 华东师范大学 基于深度学习的嗓音异常识别和评价的方法及系统
CN115062763B (zh) * 2022-06-17 2025-09-09 云南大学 一种基于Residual Bi-LSTM网络的端到端心音分割方法
CN115067945A (zh) * 2022-08-22 2022-09-20 深圳市海清视讯科技有限公司 疲劳检测方法、装置、设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104819846A (zh) * 2015-04-10 2015-08-05 北京航空航天大学 一种基于短时傅里叶变换和稀疏层叠自动编码器的滚动轴承声音信号故障诊断方法
CN105206270A (zh) * 2015-08-20 2015-12-30 长安大学 一种组合pca和rbm的孤立数字语音识别分类系统及方法
CN106847309A (zh) * 2017-01-09 2017-06-13 华南理工大学 一种语音情感识别方法
CN107545890A (zh) * 2017-08-31 2018-01-05 桂林电子科技大学 一种声音事件识别方法

Family Cites Families (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101452698B (zh) * 2007-11-29 2011-06-22 中国科学院声学研究所 一种自动嗓音谐噪比分析方法
CN103258545A (zh) * 2012-12-20 2013-08-21 苏州大学 一种病理嗓音细分方法
US10360901B2 (en) * 2013-12-06 2019-07-23 Nuance Communications, Inc. Learning front-end speech recognition parameters within neural network training
CN103778913A (zh) * 2014-01-22 2014-05-07 苏州大学 一种病理嗓音的识别方法
US9949714B2 (en) * 2015-07-29 2018-04-24 Htc Corporation Method, electronic apparatus, and computer readable medium of constructing classifier for disease detection
KR102313028B1 (ko) * 2015-10-29 2021-10-13 삼성에스디에스 주식회사 음성 인식 시스템 및 방법
KR102033411B1 (ko) * 2016-08-12 2019-10-17 한국전자통신연구원 주의집중 기반 문맥 종속 음향 모델을 이용한 음성 인식 장치 및 방법
CN106328150B (zh) * 2016-08-18 2019-08-02 北京易迈医疗科技有限公司 嘈杂环境下的肠鸣音检测方法、装置及系统
CN106901758B (zh) * 2017-02-23 2019-10-25 南京工程学院 一种基于卷积神经网络的言语置信度评测方法
CN107068167A (zh) * 2017-03-13 2017-08-18 广东顺德中山大学卡内基梅隆大学国际联合研究院 融合多种端到端神经网络结构的说话人感冒症状识别方法
CN106821337B (zh) * 2017-04-13 2020-01-03 南京理工大学 一种有监督的鼾声来源识别方法
CN107221326B (zh) * 2017-05-16 2021-05-28 百度在线网络技术（北京）有限公司 基于人工智能的语音唤醒方法、装置和计算机设备
CN107529645B (zh) * 2017-06-29 2019-09-10 重庆邮电大学 一种基于深度学习的心音智能诊断系统及方法
CN107423398B (zh) * 2017-07-26 2023-04-18 腾讯科技（上海）有限公司 交互方法、装置、存储介质和计算机设备
CN107928673B (zh) * 2017-11-06 2022-03-29 腾讯科技（深圳）有限公司 音频信号处理方法、装置、存储介质和计算机设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104819846A (zh) * 2015-04-10 2015-08-05 北京航空航天大学 一种基于短时傅里叶变换和稀疏层叠自动编码器的滚动轴承声音信号故障诊断方法
CN105206270A (zh) * 2015-08-20 2015-12-30 长安大学 一种组合pca和rbm的孤立数字语音识别分类系统及方法
CN106847309A (zh) * 2017-01-09 2017-06-13 华南理工大学 一种语音情感识别方法
CN107545890A (zh) * 2017-08-31 2018-01-05 桂林电子科技大学 一种声音事件识别方法

Also Published As

Publication number Publication date
CN109192226A (zh) 2019-01-11

Similar Documents

Publication Publication Date Title
Mouawad et al. Robust detection of COVID-19 in cough sounds: using recurrence dynamics and variable Markov model
JP7850135B2 (ja) 呼吸器症候群を検出するためのアンサンブル機械学習モデル
CN107492382B (zh) 基于神经网络的声纹信息提取方法及装置
WO2020000523A1 (zh) 一种信号处理方法及装置
CN109686377B (zh) 音频识别方法及装置、计算机可读存储介质
JP4881980B2 (ja) 音声及び映像に基づく性別−年齢識別方法及びその装置
WO2021208287A1 (zh) 用于情绪识别的语音端点检测方法、装置、电子设备及存储介质
CN116649980B (zh) 基于人工智能的情绪监测方法、系统、设备及存储介质
CN111938650B (zh) 监测睡眠时呼吸暂停的方法及装置
CN113409771B (zh) 一种伪造音频的检测方法及其检测系统和存储介质
CN113674767A (zh) 一种基于多模态融合的抑郁状态识别方法
CN110755108A (zh) 一种基于智能听诊器的心脏声音分类方法、系统、装置及可读存储介质
Wang et al. Automatic hypernasality detection in cleft palate speech using CNN
US12484555B2 (en) Poultry voiceprint identification method and system
Deb et al. Detection of common cold from speech signals using deep neural network
Nguyen et al. Fruit-CoV: An efficient vision-based framework for speedy detection and diagnosis of SARS-CoV-2 infections through recorded cough sounds
CN115273908B (zh) 一种基于分类器融合的生猪咳嗽声音识别方法
CN113724731A (zh) 利用音频判别模型进行音频判别的方法和装置
Patel et al. Multi feature fusion for COPD classification using deep learning algorithms
CN115363621A (zh) 一种肺音智能在线监测系统
Sharan et al. Detecting cough recordings in crowdsourced data using CNN-RNN
CN116246658A (zh) 一种基于MobileCBAM-CenterNet的鼾声检测方法及装置
CN119226959B (zh) 一种基于非接触式传感器的鼾声识别方法及系统
CN118711608A (zh) 脑卒中后构音障碍分类模型训练方法、分类方法及装置
CN120408360A (zh) 基于时频特征融合的双通道心肺音智能分析方法

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: 18924170

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

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

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 08.04.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18924170

Country of ref document: EP

Kind code of ref document: A1