WO2020140374A1 - 语音数据处理方法、装置、设备及存储介质 - Google Patents

语音数据处理方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2020140374A1
WO2020140374A1 PCT/CN2019/088976 CN2019088976W WO2020140374A1 WO 2020140374 A1 WO2020140374 A1 WO 2020140374A1 CN 2019088976 W CN2019088976 W CN 2019088976W WO 2020140374 A1 WO2020140374 A1 WO 2020140374A1
Authority
WO
WIPO (PCT)
Prior art keywords
voice data
sample
data segment
label
voice
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/088976
Other languages
English (en)
French (fr)
Inventor
刘博卿
贾雪丽
王健宗
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2020140374A1 publication Critical patent/WO2020140374A1/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
    • G10L15/00—Speech recognition
    • G10L15/04—Segmentation; Word boundary detection
    • 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
    • G10L15/00—Speech recognition
    • G10L15/08—Speech classification or search
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04M—TELEPHONIC COMMUNICATION
    • H04M3/00—Automatic or semi-automatic exchanges
    • H04M3/42—Systems providing special services or facilities to subscribers
    • H04M3/50—Centralised arrangements for answering calls; Centralised arrangements for recording messages for absent or busy subscribers ; Centralised arrangements for recording messages
    • H04M3/51—Centralised call answering arrangements requiring operator intervention, e.g. call or contact centers for telemarketing
    • H04M3/5166—Centralised call answering arrangements requiring operator intervention, e.g. call or contact centers for telemarketing in combination with interactive voice response systems or voice portals, e.g. as front-ends

Definitions

  • the present application relates to the field of computer technology, and in particular, to a voice data processing method, device, equipment, and storage medium.
  • Voice segmentation refers to: segmenting each person's speech segment by acquiring the speech boundary points of each two people (the boundary points may refer to the transition points where each two people speak). In practice, it is necessary to manually analyze the speech to obtain the speech boundary points of every two people, resulting in low efficiency and accuracy of speech segmentation.
  • Embodiments of the present application provide a voice data processing method, device, equipment, and storage medium, which can automatically detect the boundary points of voice data, and improve the efficiency and accuracy of voice segmentation.
  • an embodiment of the present application provides a voice data processing method, including:
  • voice data to be processed which is composed of voice data segments of multiple objects
  • the preset processing rule includes a data filtering rule and/or a data format processing rule
  • a boundary point of the target voice data is determined according to the predicted label of each voice data segment, so as to segment the voice data segment of each object from the target voice data according to the boundary point.
  • an embodiment of the present application provides a voice data processing apparatus, including:
  • An obtaining unit configured to obtain voice data to be processed, the voice data to be processed is composed of voice data segments of multiple objects;
  • a processing unit configured to process the voice data to be processed according to a preset processing rule to obtain target voice data, and the preset processing rule includes a data filtering rule and/or a data format processing rule;
  • a prediction unit configured to divide the target voice data to obtain multiple voice data segments; input the multiple voice data segments into a voice network model for prediction, to obtain a prediction tag for each voice data segment,
  • the prediction label includes the probability that the voice data segment is a boundary point;
  • a segmentation unit is configured to determine a boundary point of the target voice data according to the prediction label of each voice data segment, so as to segment the voice data segment of each object from the target voice data according to the boundary point.
  • embodiments of the present application provide another electronic device, including:
  • a processor suitable for implementing one or more instructions
  • a computer-readable storage medium stores one or more instructions, the one or more instructions are adapted to be loaded by the processor and perform the following steps:
  • voice data to be processed which is composed of voice data segments of multiple objects
  • the preset processing rule includes a data filtering rule and/or a data format processing rule
  • a boundary point of the target voice data is determined according to the predicted label of each voice data segment, so as to segment the voice data segment of each object from the target voice data according to the boundary point.
  • an embodiment of the present application provides a computer-readable storage medium, including: the computer-readable storage medium stores one or more instructions, the one or more instructions are suitable for being loaded and executed by a processor The following steps:
  • voice data to be processed which is composed of voice data segments of multiple objects
  • the preset processing rule includes a data filtering rule and/or a data format processing rule
  • a boundary point of the target voice data is determined according to the predicted label of each voice data segment, so as to segment the voice data segment of each object from the target voice data according to the boundary point.
  • the boundary point of the target voice data can be automatically obtained through the voice network model, no manual operation is required, a lot of manpower can be saved, and the accuracy and efficiency of obtaining the boundary point of the voice data can be improved; Processing of intelligent, automated needs.
  • FIG. 1 is a schematic flowchart of a voice data processing method provided by an embodiment of the present application.
  • FIG. 2 is a schematic flowchart of a voice data processing method provided by an embodiment of the present application.
  • FIG. 3 is a schematic structural diagram of a voice data processing device according to an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • the embodiments of the present application provide an automatic voice data processing method, which can be performed by an electronic device
  • the electronic device may refer to devices such as smart terminals, servers, computers, or detectors; this method can predict the voice data through the voice network model to obtain the predicted label, determine the boundary point of the voice data according to the predicted label, and determine The boundary points of each segment segment the voice data segment of each object, which can save a lot of manpower, and can improve the detection accuracy, to meet the user's voice segmentation intelligence and automation needs.
  • FIG. 1 is a schematic flowchart of a voice data processing method provided in an embodiment of the present application.
  • the method in the embodiment of the present application may be executed by the electronic device mentioned above.
  • the voice data processing method includes the following steps.
  • the specific content included in the processed voice data under different application scenarios is different.
  • the to-be-processed voice data may refer to voice data obtained by recording words spoken by multiple people; in a call application scenario, the to-be-processed voice data may refer to multiple calls received by the call center.
  • Personal call data The object here mainly refers to the person speaking, and the object can also refer to an animal.
  • the preset processing rule includes a data filtering rule and/or a data format processing rule.
  • the voice data to be processed includes redundant data, such as mute data or non-human voice, etc.
  • the non-human voice includes device sounds (such as camera sounds) and environmental sounds (such as vehicle sounds) );
  • the voice data to be processed can be pre-processed.
  • the electronic device may use a filter to filter the speech data to be processed to obtain target speech data, where the filter may be a high-pass filter or a band-pass filter.
  • the electronic device may perform format conversion processing on the voice data to be processed.
  • the preset processing rule includes a data filtering rule
  • step S102 includes: dividing the voice data to be processed to obtain multiple original voice data segments, and acquiring each of the multiple original voice data segments The energy values of the original speech data segments, delete the original speech data segments whose energy values in the plurality of original speech data segments are less than or equal to the preset energy value, and the energy values in the plurality of original speech data segments are greater than the preset energy
  • the original voice data segments of the value are combined to obtain the target voice data.
  • the electronic device can filter the voice data to be processed according to the energy of the voice data.
  • the electronic device may divide the voice data to be processed according to a preset length to obtain multiple original voice data segments, and perform time-frequency transformation on each of the multiple original voice data segments to obtain each Frequency domain information of each original voice data segment.
  • the frequency domain information of each original voice data segment is used to describe the relationship between the frequency and energy of each original voice data segment. According to the frequency domain information of each original voice data segment, the energy value of the corresponding original voice data segment is obtained.
  • the original voice data segment with the smaller energy value has a higher probability of interfering with the voice data
  • the original voice data segment with the higher energy value is The probability of useful voice data is high; therefore, the original voice data segments with energy values less than or equal to the preset energy value in the multiple original voice data segments can be deleted, and the energy values in the multiple original voice data segments are greater than the preset
  • the original voice data segments of the energy value are combined to obtain the target voice data.
  • the preset processing rule includes a data format processing rule
  • step S102 includes: acquiring the data format of the voice data to be processed, when the data format of the voice data to be processed is different from the preset data format When they are the same, format conversion processing is performed on the to-be-processed voice data according to the preset data format to obtain the target voice data.
  • the electronic device can obtain the data format of the voice data to be processed.
  • the data format of the voice data to be processed is different from the preset data format, it indicates that the voice data to be processed
  • the data format is not suitable for voice network model prediction processing.
  • the preset data format can be a data format suitable for voice network model prediction of voice data.
  • the format can be Pulse Code Modulation (Pulse Code Modulation, PCM), or voice Interactive format (Audio Interchange File Format, AIFF), etc.
  • PCM Pulse Code Modulation
  • AIFF Audio Interchange File Format
  • S103 Divide the target voice data to obtain multiple voice data segments, and input the multiple voice data segments into the voice network model for prediction, to obtain a prediction tag for each voice data segment, where the prediction tag includes the voice data segment Is the probability of the boundary point.
  • the electronic device may divide the target voice data into multiple voice data segments that do not overlap with each other, or, to improve the accuracy of prediction, the electronic device may use the target voice
  • the data is divided into multiple voice data segments that overlap each other.
  • a plurality of voice data segments are input into the voice network model, and the voice network model analyzes and predicts the feature points in each voice data segment to obtain the predicted label of each voice data segment.
  • the electronic device may divide the target voice data into multiple voice data segments that do not overlap with each other. Specifically, the electronic device may divide multiple voice data segments into multiple voice data segments according to a preset voice length. For example, the preset voice length is 10s, and the length of the voice data to be processed is 100s. The electronic device may divide the voice data to be processed into 10 voice data segments, for example, the first voice data segment is 0-10s, The second voice data segment is 10-20s, the third voice data segment is 20s-30s, and so on to obtain 10 voice data segments.
  • each voice data segment is a non-overlapping voice data segment, because the speech transitions between various objects are fast, it is easy to cause no boundary point to be detected.
  • the first voice data segment and the second voice data segment are non-overlapping and adjacent voice data segments. If the first voice data segment is the voice data of the object A, if the second voice data segment is the voice data of the object B. Since the feature points in the first voice data are similar and the feature points in the second voice data are similar, neither the first voice data nor the second voice data will be predicted as boundary points, resulting in undetectable Boundary point. To avoid missing the boundary point, the electronic device may divide the target voice data into multiple voice data segments that overlap each other.
  • the target voice data is divided according to a preset step size to obtain multiple voice data segments, and the length of each voice data segment in the multiple voice data segments is greater than the preset step size.
  • the length of each voice data segment may be the same or different.
  • the following takes the length of each voice data segment as an example.
  • the preset step size is 8s, and the length of each voice data is 10s.
  • One voice data segment is 0-10s, the second voice data segment is 8-18s, and the third voice data segment is 16-26s. Every two adjacent voice data segments include overlapping voice data.
  • the foregoing inputting the plurality of voice data segments into the voice network model for prediction to obtain a prediction tag for each voice data segment includes: the voice network model analyzing the similarity of each feature point in each voice data segment Degree, according to the similarity of each feature point, calculate the sum of the similarity of each voice data segment, and determine the predicted label of the corresponding voice data segment according to the sum of the similarity of each voice data segment.
  • the voice data segment includes 5 feature points, respectively x1, x2, x3, x4, and x5.
  • the electronic device can separately calculate the similarity between each feature point and other feature points, for example, calculate the similarity between x1 and x2, x3, x4, and x5, and accumulate the similarities to obtain the total similarity of x1.
  • the sum of similarities of x2-x5 is calculated by calculation, the sum of similarities of each feature point is accumulated to obtain the sum of similarities of the first voice data segment, and the first voice data is determined according to the sum of similarities of the first voice data segment
  • the prediction tag includes the probability that the voice data segment is a boundary point. The value of this probability is [0,1]. The larger the probability value, the probability that the corresponding voice data segment (or feature point in the voice data segment) is a boundary point.
  • the probability value is 0, the corresponding voice data segment (or feature point in the voice data segment) is not a boundary point; if the probability value is 1, it corresponds to the voice data segment (or feature in the voice data segment Point) is the boundary point.
  • the feature points of the voice data segment may refer to any one or more of the following energy, pitch and timbre, etc.
  • Energy refers to the intensity of the sound (ie, the loudness of the sound)
  • pitch refers to the level of the sound
  • Timbre refers to the characteristics of sound.
  • the above calculation of the similarity sum of each voice data segment includes: calculating the corresponding voice data according to the similarity between the energy feature points of each voice data segment
  • the sum of the first similarity of the segments, the sum of the second similarity of the corresponding voice data segments is calculated according to the similarity between the tone feature points of each voice data segment, and the similarity between the timbre feature points of each voice data segment
  • the third similarity sum of the corresponding speech data segment is calculated by calculating the degree; the first similarity sum, the second similarity sum and the third similarity sum are weighted and summed to obtain the similarity sum of each voice data segment.
  • the weights corresponding to the sum of the first similarity, the sum of the second similarity, and the sum of the third similarity may be set by the user, or may be set by the electronic device according to the application scenario.
  • the weight of the sum of the first similarity may be set to a larger value to highlight the difference in sound energy of each object.
  • the weight of the sum of the second similarity may be set to a large value to highlight the difference in pitch of each object.
  • the weight of the sum of the third similarity may be set to a larger value to highlight the difference in timbre of each object.
  • the speech network model may be composed of two long-short-term memory network models (ie, Bi-LSTMs) and a multi-layer neural network model, and the multi-layer neural network model may be connected to one of the long-short memory network models.
  • Each Bi-LSTMs includes forward long-short-term memory network layer forward LSTM and backward long-short-term memory network layer backward LSTM, the outputs of the two are connected together, and then output to the next layer; so that Bi-LSTMs can pass forward and Reverse the two directions to process the voice data segment, so that the previous and future context information can be used at the same time.
  • the multi-layer neural network model consists of three fully connected layers.
  • the activation equation of the first two layers can be the tanh function, and the activation equation of the last layer can be the igmoid function; so that a probability between 0 and 1 can be output (prediction label).
  • S104 Determine a boundary point of the target voice data according to the predicted label of each voice data segment, so as to segment the voice data segment of each object from the target voice data according to the boundary point.
  • the electronic device may determine the boundary point of the target voice data according to the prediction tag of the voice data segment. For example, the electronic device may divide the voice data segment with a probability greater than a preset probability As a boundary point of the target voice data, or a part of feature points in the voice data segment with a probability greater than a preset probability as the boundary point of the target voice data. And segment the speech data segment of each object from the target speech data according to the boundary edge, so as to analyze the speech data segment of a certain object.
  • the boundary point of the target voice data may refer to a conversion point of voice data segments of different objects.
  • the first voice data segment to the fourth voice data segment are the first The voice data segment of the object (such as the first person), and the voice data segment after the fourth voice data segment are the voice data segment of the second object (such as the second person).
  • the boundary point of the target voice data can be automatically obtained through the voice network model, no manual operation is required, a lot of manpower can be saved, and the accuracy and efficiency of obtaining the boundary point of the voice data can be improved; Processing of intelligent, automated needs.
  • FIG. 2 is a schematic flowchart of another voice data processing method provided in an embodiment of the present application.
  • the method in the embodiment of the present application may be executed by the electronic device mentioned above.
  • the voice data processing method includes the following steps.
  • the training sample set includes a plurality of sample speech data segments of sample audio data, and a label tag of each sample speech data segment, where the label tag includes the probability that the sample speech data segment is a boundary point.
  • the voice network model may be optimally trained.
  • the electronic device may collect a training sample set, where the training sample set includes a plurality of sample speech data segments of sample audio data, and a label of each sample speech data segment.
  • the labeling tag includes the probability that the sample speech data segment is a boundary point, and the labeling tag may refer to manually labeling the feature sample.
  • the sample audio data may be composed of voice data segments of objects in different regions, or/and, the sample audio data may be composed of voice data segments of objects of different ages.
  • the plurality of sample voice data segments are used as the input of the voice network model, and the label of each sample voice data segment is used as the training target of the voice network model; when the voice network model outputs the prediction of the voice data segment When the label is the same as the label of the voice data segment, or the two are similar, it indicates that the prediction accuracy of the voice network model is high, and then the iterative training of the voice network model is ended; when the voice data segment output by the voice network model The difference between the predicted label and the labeled label of the voice data segment is large, indicating that the prediction accuracy of the voice network model is low, then the network parameters of the voice network model are adjusted, and iterative training of the voice network model is continued.
  • step S202 includes: inputting the plurality of sample speech data segments into the speech network model for prediction to obtain a predicted label for each sample speech data segment; according to the predicted label for each sample speech data segment and The prediction label of the corresponding sample voice data segment determines the prediction error of the voice network model; if the prediction error of the voice network model is greater than the preset error value, the network parameters of the voice network model are adjusted; if the prediction error of the voice network model Less than or equal to the preset error value, the iterative training of the voice network model is ended.
  • the electronic device can optimize the voice network model by adjusting the network parameters of the voice network model.
  • the multiple sample voice data segments are input into the voice network model for prediction, and the predicted label of each sample voice data segment is obtained.
  • the prediction label of each sample voice data segment and the corresponding label of the sample voice data segment determine the prediction error of the voice network model; if the prediction error of the voice network model is greater than the preset error value, it indicates the prediction accuracy of the voice network model Low, adjust the network parameters of the voice network model, and continue to iteratively train the voice network model; if the prediction error of the voice network model is less than or equal to the preset error value, it indicates that the prediction accuracy of the voice network model is High, you can end the iterative training of the voice network model.
  • the labeled label of the data segment is y i
  • the predicted label of the i-th speech data segment is f(X) i
  • the prediction error L of the speech network model can be expressed as formula (1).
  • the first type label tag is a label whose sample voice data segment is a boundary point with a probability greater than a preset probability
  • the target sample voice data segment is the voice data segment where the boundary point of the sample audio data is located
  • the label label of the sample voice data segment with a time interval less than or equal to the preset time interval from the target sample voice data segment is set to the first Class labeling tag
  • the second type labeling tag is the sample speech data segment as The label with the probability of the boundary point is less than or equal to the preset probability.
  • the voice data includes boundary points and non-boundary points. Because there are relatively few boundary points in the voice data, that is, the probability that only a few voice data segments are boundary points is greater than the preset threshold, and the probability that most voice data segments are boundary points Less than the preset threshold; the imbalance (too much difference) between the two types of data causes many problems when training the voice network model, such as the prediction accuracy of the voice network model is low. Therefore, it is possible to increase the number of positive samples near the true boundary point, that is, increase the voice data segment whose probability is greater than the preset probability. Specifically, the electronic device may receive an instruction to set the label of the target sample voice data segment to the first type of label tag.
  • the target sample voice data segment is the voice data segment where the boundary point of the sample audio data is located, that is, the target sample voice data
  • the segment is the voice data segment where the real boundary point is located.
  • the label label of the sample speech data segment with a time interval less than or equal to the preset time interval from the target sample speech data segment can be set as the first type label label, that is, near the target sample speech data segment
  • the voice data segments of are marked as the voice data segment where the boundary point (ie positive sample) is located.
  • the sample speech data segment at the time interval is marked as the speech data segment where the non-boundary point is located.
  • the preset time interval is 50 ms, and the electronic device may mark all sample speech data segments within a time interval of 50 ms from the target sample data segment as the first type of labeling tags.
  • the preset processing rule includes a data filtering rule and/or a data format processing rule.
  • each voice data segment includes multiple feature points, and each adjacent two voice data segments include the same feature point, and the predicted label of each voice data segment is each feature in the corresponding voice data segment
  • the predicted label of the point, step S206 includes: counting the average probability that each feature point is a boundary point according to the predicted label of each speech data segment, and taking all feature points whose average probability is greater than a preset probability as the boundary point of the target speech data .
  • each adjacent voice data segment has overlap, each adjacent two voice data segments include the same feature point.
  • the electronic device may The average probability to determine the boundary point of voice data. For example, if the target voice data includes a first voice data segment and a second voice data segment, the first voice data segment includes feature points x1, x2, x3, x4, and x5, and the predicted label of the first voice data segment is 0.2, then the feature The predicted labels (probabilities) of points x1, x2, x3, x4 and x5 are all 0.2; the second speech data segment includes feature points x4, x5, x6, x7 and x8.
  • the predicted label of the second speech data segment is 0.5, then the predicted labels of the feature points x4, x5, x6, x7, and x8 are all 0.5; the average label (ie, the average probability) of each feature point is determined according to the predicted label, that is, the feature point
  • the average predicted labels of x1, x2, and x3 are all 0.2, the average predicted labels of feature points x4 and x5 are 0.35, and the average predicted labels of feature points x6, x7, and x8 need to be calculated according to the predicted labels of the third speech data segment. Further, all feature points whose average probability is greater than a preset probability are taken as the boundary points of the target voice data.
  • the boundary point of the target voice data can be automatically obtained through the voice network model, no manual operation is required, a lot of manpower can be saved, and the accuracy and efficiency of obtaining the boundary point of the voice data can be improved; Processing of intelligent, automated needs.
  • FIG. 3 is a schematic structural diagram of a voice data processing apparatus provided by an embodiment of the present application.
  • the apparatus of the embodiment of the present application may be provided in the electronic device mentioned above.
  • the device includes:
  • the obtaining unit 301 is configured to obtain voice data to be processed, the voice data to be processed is composed of voice data segments of multiple objects.
  • the processing unit 302 is configured to process the voice data to be processed according to a preset processing rule to obtain target voice data.
  • the preset processing rule includes a data filtering rule and/or a data format processing rule.
  • the prediction unit 303 is used to divide the target voice data to obtain multiple voice data segments; input the multiple voice data segments into a voice network model for prediction, and obtain a prediction tag for each voice data segment.
  • the prediction tag includes the probability that the voice data segment is a boundary point.
  • the segmentation unit 304 is configured to determine a boundary point of the target voice data according to the prediction tag of each voice data segment, so as to segment the voice data segment of each object from the target voice data according to the boundary point.
  • the processing unit 302 is configured to divide the target voice data according to a preset step size to obtain multiple voice data segments, and the length of each voice data segment in the multiple voice data segments is greater than the Set step size.
  • the training unit 305 is configured to obtain a set of training samples, the set of training samples includes a plurality of sample speech data segments of sample audio data, and a label tag for each sample speech data segment, the label tag includes sample speech The probability of a data segment being a boundary point; taking the multiple sample voice data segments as input to the voice network model, and using the label of each sample voice data segment as the training target of the voice network model
  • the voice network module performs iterative training.
  • the training unit 305 is configured to input the plurality of sample voice data segments into the voice network model for prediction, to obtain a predicted label for each sample voice data segment; according to the The prediction label and the corresponding label of the sample voice data segment determine the prediction error of the voice network model; if the prediction error of the voice network model is greater than a preset error value, adjust the network parameters of the voice network model; If the prediction error of the voice network model is less than or equal to the preset error value, the iterative training of the voice network model ends.
  • the training unit 305 is configured to receive an instruction to set the label tag of the target sample voice data segment as the first type label tag, where the probability that the first type label tag is the sample voice data segment as the boundary point is greater than the preset Probability label, the target sample voice data segment is the voice data segment where the boundary point of the sample audio data is located; the sample voice data segment whose time interval from the target sample voice data segment is less than or equal to the preset time interval Is set as the first type label tag; the label tag of the sample speech data segment with a time interval greater than the preset time interval from the target sample speech data segment is set as the second type label tag, the second The class label is a label whose probability that the sample speech data segment is a boundary point is less than or equal to the preset probability.
  • each voice data segment includes multiple feature points, and each adjacent two voice data segments include the same feature point, and the predicted label of each voice data segment is each feature in the corresponding voice data segment Predictive tags of points; a segmentation unit 304 for counting the average probability that each feature point is a boundary point according to the predictive tags of each voice data segment; taking all feature points whose average probability is greater than a preset probability as the target speech The boundary point of the data.
  • the preset processing rule includes a data filtering rule
  • the processing unit 302 is configured to divide the voice data to be processed to obtain multiple original voice data segments; and obtain the multiple original voice data segments The energy value of each original voice data segment in the; delete the original voice data segments whose energy value is less than or equal to the preset energy value in the plurality of original voice data segments; delete the energy values in the plurality of original voice data segments greater than The original voice data segments of the preset energy value are combined to obtain the target voice data.
  • the boundary point of the target voice data can be automatically obtained through the voice network model, no manual operation is required, a lot of manpower can be saved, and the accuracy and efficiency of obtaining the boundary point of the voice data can be improved; Processing of intelligent, automated needs.
  • FIG. 4 is a schematic structural diagram of an electronic device according to an embodiment of the present application.
  • the electronic device in this embodiment may include: one or more processors 401; one or more input devices 402 , One or more output devices 403 and memory 404.
  • the processor 401, the input device 402, the output device 403, and the memory 404 are connected via a bus 404.
  • the processor 401 may be a central processing unit (Central Processing Unit, CPU), the processor may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC ), ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the input device 402 may include a touchpad, a fingerprint sensor (for collecting user's fingerprint information and fingerprint direction information), a microphone, etc.
  • the output device 403 may include a display (LCD, etc.), a speaker, etc.
  • the output device 403 may output a prompt Information, prompt information can be used to prompt the boundary point of the target voice data.
  • the memory 404 may include a read-only memory and a random access memory, and provide instructions and data to the processor 401.
  • a part of the memory 404 may further include a non-volatile random access memory, the memory 404 is used to store a computer program, the computer program includes program instructions, and the processor 401 is used to execute the program instructions stored in the memory 404, for performing a Voice data processing method, which is used to perform the following operations:
  • voice data to be processed which is composed of voice data segments of multiple objects
  • the preset processing rule includes a data filtering rule and/or a data format processing rule
  • a boundary point of the target voice data is determined according to the predicted label of each voice data segment, so as to segment the voice data segment of each object from the target voice data according to the boundary point.
  • the processor 401 is used to execute the program instructions stored in the memory 404 to perform the following operations:
  • the processor 401 is used to execute the program instructions stored in the memory 404 to perform the following operations:
  • the training sample set including a plurality of sample speech data segments of sample audio data, and a labeling tag of each sample speech data segment, the labeling tag including the probability that the sample speech data segment is a boundary point;
  • the plurality of sample voice data segments are used as input to the voice network model, and the label of each sample voice data segment is used as the training target of the voice network model to iteratively train the voice network module.
  • the processor 401 is used to execute the program instructions stored in the memory 404 to perform the following operations:
  • the iterative training of the voice network model is ended.
  • the processor 401 is used to execute the program instructions stored in the memory 404 to perform the following operations:
  • the first type label tag is a label whose sample voice data segment is a boundary point with a probability greater than a preset probability, and the target sample voice data The segment is the voice data segment where the boundary point of the sample audio data is located;
  • the label tag of the sample speech data segment with a time interval greater than the preset time interval from the target sample speech data segment is set as a second type label tag, where the second type label tag is the sample speech data segment as a boundary point With a probability less than or equal to the preset probability.
  • the processor 401 is used to execute the program instructions stored in the memory 404 to perform the following operations:
  • All feature points with an average probability greater than a preset probability are used as boundary points of the target voice data.
  • the processor 401 is used to execute the program instructions stored in the memory 404 to perform the following operations:
  • the processor 401, the input device 402, and the output device 403 described in the embodiments of the present application may execute the implementation methods described in the first and second embodiments of the voice data processing method provided by the embodiments of the present application, or The implementation of the electronic device described in the embodiments of the present application will not be repeated here.
  • a computer-readable storage medium is also provided in an embodiment of the present application.
  • the computer-readable storage medium stores a computer program, and the computer program includes program instructions.
  • FIG. 1 and FIG. 2 The voice data processing method shown in the embodiment.
  • the computer-readable storage medium may be an internal storage unit of the medical management device according to any of the foregoing embodiments, such as a hard disk or a memory of the control device.
  • the computer-readable storage medium may also be an external storage device of the control device, for example, a plug-in hard disk equipped on the control device, a smart memory card (Smart) Media (SMC), a secure digital (SD) card, Flash card (Flash Card), etc.
  • the computer-readable storage medium may also include both an internal storage unit of the control device and an external storage device.
  • the computer-readable storage medium is used to store the computer program and other programs and data required by the control device.
  • the computer-readable storage medium can also be used to temporarily store data that has been or will be output.
  • control device and method may be implemented in other ways.
  • the device embodiments described above are schematic.
  • the division of the unit may be a logical function division, and there may be other divisions in actual implementation, for example, multiple units or components may be combined or integrated To another system, or some features can be ignored, or not implemented.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Telephonic Communication Services (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种语音数据处理方法、装置、设备及储存介质，该方法包括：获取待处理的语音数据，待处理的语音数据由多个对象的语音数据段组成（S101）；按照预设的处理规则对待处理的语音数据进行处理，得到目标语音数据（S102）；对目标语音数据进行划分，得到多个语音数据段；将多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签（S103）；根据每个语音数据段的预测标签确定目标语音数据的边界点，根据边界点从目标语音数据中分割出每个对象的语音数据段（S104）。该方法能够自动获取语音数据的边界点，并且能够提高获取语音数据的边界点的准确度。

Description

语音数据处理方法、装置、设备及存储介质
本申请要求于2019年01月04日提交中国专利局、申请号为201910018423.9、申请名称为“语音数据处理方法、装置、设备及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，尤其涉及一种语音数据处理方法、装置、设备及存储介质。
背景技术
通常呼叫中心接收到的语音数据很多都混杂有多人的语音片段，这时需要先对语音数据进行语音分割(speaker diarization)，才能进一步对目标语音片段进行语音分析。语音分割是指：通过获取每两个人的语音边界点(边界点可以是指每两个人说话的转换点)来分割出每个人的语音段。实践中，需要人工对语音进行分析，以获取每两个人的语音边界点，导致语音分割的效率及准确度较低。
发明内容
本申请实施例提供一种语音数据处理方法、装置、设备及存储介质，可自动检测出语音数据的边界点，提高语音分割的效率及准确度。
第一方面，本申请实施例提供了一种语音数据处理方法，包括：
获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
按照预设的处理规则对所述待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
第二方面，本申请实施例提供了一种语音数据处理装置，包括：
获取单元，用于获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
处理单元，用于按照预设的处理规则对待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
预测单元，用于对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
分割单元，用于根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
第三方面，本申请实施例提供了另一种电子设备，包括：
处理器，适于实现一条或一条以上指令；以及，
计算机可读存储介质，所述计算机可读存储介质存储有一条或一条以上指令，所述一条或一条以上指令适于由所述处理器加载并执行如下步骤：
获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
按照预设的处理规则对所述待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
第四方面，本申请实施例提供了一种计算机可读存储介质，包括：所述计算机可读存储介质存储有一条或一条以上指令，所述一条或一条以上指令适于由处理器加载并执行如下步骤：
获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
按照预设的处理规则对所述待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
本申请实施例中，可通过语音网络模型自动获取目标语音数据的边界点，不需要人工操作，可节省大量人力，并可提高获取语音数据的边界点的准确度及效率；满足用户对语音数据处理的智能化、自动化需求。
附图说明
图1是本申请实施例提供的一种语音数据处理方法的流程示意图；
图2是本申请实施例提供的一种语音数据处理方法的流程示意图；
图3是本申请实施例提供的一种语音数据处理装置的结构示意图；
图4是本申请实施例提供的一种电子设备的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请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申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，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，都属于本申请保护的范围。
基于现有技术中需要人工分析方式确定语音数据的边界点，导致语音分割的效率及准确度较低的问题，本申请实施例提供一种自动语音数据处理的方法，该方法可以由电子设备来执行，该电子设备可以是指智能终端、服务器、电脑或检测仪等设备；该方法可以通过语音网络模型对语音数据进行预测，得到预测标签，根据预测标签确定语音数据的边界点，根据语音数据的边界点分割出每个对象的语音数据段，可节省大量人力，并可提高检测准确度，满足用户语音分割智能化、 自动化需求。
请参见图1，是本申请实施例提供的一种语音数据处理方法的流程示意图，本申请实施例的所述方法可以由上述提及的电子设备来执行。本实施例中，该语音数据处理方法包括以下步骤。
S101、获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成。
本申请实施例中，该处理的语音数据在不同应用场景下所包括的具体内容不同。例如，在会议场景中，该待处理的语音数据可以是指对多个人所讲的话进行录音得到的语音数据；在呼叫应用场景中，该待处理的语音数据可以是指呼叫中心接收到的多个人的呼叫数据。这里的对象主要是指说话的人，对象也可以是指动物。
S102、按照预设的处理规则对待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则。
本申请实施例中，由于待处理的语音数据中包括冗余数据，如静音数据或非人的语音等等，非人的语音包括设备的声音(如拍照声)、环境声音(如车辆的声音)；为了避免对冗余数据进行处理，提高处理效率，可以对待处理的语音数据进行预处理。具体的，电子设备可以采用滤波器对待处理的语音数据进行过滤处理，得到目标语音数据，其中，滤波器可以为高通滤波器或带通滤波器等。和/或者，为了便于语音网络模型对语音数据进行预测，电子设备可以对待处理的语音数据进行格式转换处理。
在一个实施例中，所述预设的处理规则包括数据过滤规则，步骤S102包括：将该待处理的语音数据进行划分，得到多个原始语音数据段，获取该多个原始语音数据段中每个原始语音数据段的能量值，将该多个原始语音数据段中能量值小于或等于预设能量值的原始语音数据段删除，将该多个原始语音数据段中能量值大于该预设能量值的原始语音数据段进行合并，得到该目标语音数据。
通常有用语音数据的能量大于干扰语音数据(如静音或非人的声语音)的能量，因此，电子设备可以根据语音数据的能量对待处理的语音数据进行过滤处理。具体的，电子设备可以按照预设长度将该待处理的语音数据进行划分，得到多个原始语音数据段，对该多个原始语音数据段中每个原始语音数据段进行时频变换，得到每个原始语音数据段的频域信息，每个原始语音数据段的频域信息用于描述每个原始语音数据段的频率与能量之间的关系。根据每个原始语音数据段的频域信息获取对应原始语音数据段的能量值，由于能量值较小的原始语音数据段为干扰语音数据的概率较高，能量值较高的原始语音数据段为有用语音数据的概率较高；因此可以将该多个原始语音数据段中能量值小于或等于预设能量值的原始语音数据段删除，将该多个原始语音数据段中能量值大于该预设能量值的原始语音数据段进行合并，得到该目标语音数据。
在另一个实施例中，该预设的处理规则包括数据格式处理规则，步骤S102包括：获取该待处理的语音数据的数据格式，当该待处理的语音数据的数据格式与预设数据格式不相同时，按照该预设数据格式该待处理的语音数据进行格式转换处理，得到该目标语音数据。
为了便于语音网络模型对语音数据进行预测，电子设备可以获取该待处理的语音数据的数据格式，当该待处理的语音数据的数据格式与预设数据格式不相同时，表明该待处理语音数据的数据格式不适用于语音网络模型预测处理，该预设 数据格式可以为适用于语音网络模型对语音数据预测的数据格式，例如该格式可以为脉冲编码调制(Pulse Code Modulation，PCM)，或语音交互格式(Audio Interchange File Format，AIFF)等等。按照该预设数据格式对该待处理的语音数据进行格式转换处理，得到该目标语音数据，该目标语音数据的数据格式为预设数据格式。
S103、对该目标语音数据进行划分，得到多个语音数据段，将该多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，该预测标签包括语音数据段为边界点的概率。
本申请实施例中，为了提高预测的效率，电子设备可以将该目标语音数据划分为相互之间不重叠的多个语音数据段，或者，为了提高预测的准确度，电子设备可以将该目标语音数据划分为相互之间重叠的多个语音数据段。进一步，将多个语音数据段输入到语音网络模型中，语音网络模型对每个语音数据段中的特征点进行分析并进行预测，得到每个语音数据段的预测标签。
在一个实施例中，电子设备可以将该目标语音数据划分为相互之间不重叠的多个语音数据段。具体的，电子设备可以按照预设语音长度将多个语音数据段划分为多个语音数据段。例如，该预设语音长度为10s，待处理的语音数据的长度为100s，电子设备可以将该待处理的语音数据划分为10个语音数据段，如，第一语音数据段为0-10s，第二语音数据段为10-20s，第三语音数据段为20s-30s，以此类推得到10个语音数据段。
在另一个实施例中，若各个语音数据段为不重叠的语音数据段，由于各个对象之间说话转换很快，容易导致检测不到边界点。例如，第一语音数据段和第二语音数据段为非重叠且相邻的语音数据段，若第一语音数据段为对象A的语音数据，若第二语音数据段为对象B的语音数据。由于第一语音数据内的特征点都具有相似性，第二语音数据内的特征点都具有相似性，则第一语音数据、第二语音数据都不会被预测为边界点，导致检测不到边界点。为了避免错过边界点，电子设备可以将该目标语音数据划分为相互之间重叠的多个语音数据段。具体的，按照预设步长对所述目标语音数据进行划分，得到多个语音数据段，该多个语音数据段中每个语音数据段的长度大于该预设步长。其中，每个语音数据段的长度可以相同，也可以不相同，下面以每个语音数据段的长度相同为例，例如，预设步长为8s，每个语音数据的长度为10s，则第一语音数据段为0-10s，第二语音数据段为8-18s，第三语音数据段为16-26s。每两个相邻的语音数据段之间包括重叠的语音数据。
在一个实施例中，上述将该多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，包括：语音网络模型分析每个语音数据段中各个特征点的相似度，根据各个特征点的相似度，计算每个语音数据段的相似度总和，根据每个语音数据段的相似度总和确定对应语音数据段的预测标签。其中，语音数据段的相似度总和越大，表明该语音数据段为同一个对象的语音数据的概率较大，该语音数据段为边界点的概率较小；语音数据段的相似度总和越小，表明该语音数据段为同一个对象的语音数据的概率较小，该语音数据段为边界点的概率较大。例如，对于第一语音数据段，该语音数据段包括5个特征点，分别为x1，x2，x3，x4及x5。电子设备可以分别计算每个特征点与其他特征点之间的相似度，如，分别计算x1与x2，x3，x4及x5的相似度，对各个相似度进行累加得到x1的相似度总和，同理计算得到x2-x5的相似度总和，将各个特征点的相似度总和进行累加，得到该第一语音数据段的相似度总和，根据第一语音数 据段的相似度总和确定该第一语音数据段的预测标签，并输出该预设标签。该预测标签包括语音数据段为边界点的概率，该概率的值为[0,1]，概率的值越大，表明对应语音数据段(或语音数据段内的特征点)为边界点的概率越大，概率的值越小，表明对应语音数据段(或语音数据段内的特征点)为边界点的概率越小。特别的，若概率的值为0，则对应语音数据段(或语音数据段内的特征点)不为边界点；若概率的值为1，则对应语音数据段(或语音数据段内的特征点)为边界点。
在一个实施例中，语音数据段的特征点可以是指以下任一项或多项能量、音调及音色等等，能量是指声音的强度(即声音的响度)，音调是指声音的高低，音色是指声音的特性。当语音数据段的特征点可以包括能量、音调及音色时，上述计算每个语音数据段的相似度总和，包括：根据每个语音数据段的能量特征点之间的相似度计算得到对应语音数据段的第一相似度总和，根据每个语音数据段的音调特征点之间的相似度计算得到对应语音数据段的第二相似度总和，根据每个语音数据段的音色特征点之间的相似度计算得到对应语音数据段的第三相似度总和；对第一相似度总和、第二相似度总和及第三相似度总和进行加权求和，得到每个语音数据段的相似度总和。其中，第一相似度总和、第二相似度总和及第三相似度总和对应的权重可以是用户设置的，也可以是电子设备根据应用场景设置的。例如，在各个对象的声音响度差别较大的场景中，可以将第一相似度总和的权重设置一个较大值，以突出各个对象的声音能量的差别。在各个对象的音调差别较大的场景中，可以将第二相似度总和的权重设置一个较大值，以突出各个对象的音调的差别。在各个对象的音色差别较大的场景中，可以将第三相似度总和的权重设置一个较大值，以突出各个对象的音色的差别。
在一个实施例中，该语音网络模型可以由两个长短期记忆网络模型(即Bi-LSTMs)及一个多层神经网络模型，该多层神经网络模型可以与其中一个长短期记忆网络模型连接。每个Bi-LSTMs包括前向长短期记忆网络层forward LSTM和后向长短期记忆网络层backward LSTM，两者的输出连接到一起，然后输出给下一层；以便Bi-LSTMs可以通过前向和反向两个方向来处理语音数据段，从而可以同时利用以前的和将来的上下文信息。多层神经网络模型由三个全连接层组成，前两层的激活方程可以是tanh函数，最后一层的激活方程可以s igmoid函数；以便可以输出一个范围在0和1之间的概率(预测标签)。
S104、根据该每个语音数据段的预测标签确定该目标语音数据的边界点，以根据该边界点从该目标语音数据中分割出每个对象的语音数据段。
本申请实施例中，为了分割各个对象的语音数据段，电子设备可以根据该语音数据段的预测标签确定该目标语音数据的边界点，例如，电子设备可以将概率大于预设概率的语音数据段作为目标语音数据的边界点，或将概率大于预设概率的语音数据段中的部分特征点作为目标语音数据的边界点。并根据该边界边从该目标语音数据中分割出每个对象的语音数据段，以便对某个对象的语音数据段进行分析。目标语音数据的边界点可以是指不同对象的语音数据段的转化点，例如，假设目标语音数据中第四语音数据段为边界点，则第一语音数据段到第四语音数据段为第一对象(如第一个人)的语音数据段，第四语音数据段以后的语音数据段为第二对象(如第二个人)的语音数据段。
本申请实施例中，可通过语音网络模型自动获取目标语音数据的边界点，不需要人工操作，可节省大量人力，并可提高获取语音数据的边界点的准确度及效率；满足用户对语音数据处理的智能化、自动化需求。
请参见图2，是本申请实施例提供的另一种语音数据处理方法的流程示意图，本申请实施例的该方法可以由上述提及的电子设备来执行。本实施例中，该语音数据处理方法包括以下步骤。
S201、获取训练样本集合，该训练样本集合包括样本音频数据的多个样本语音数据段，及每个样本语音数据段的标注标签，该标注标签包括样本语音数据段为边界点的概率。
S202、以该多个样本语音数据段为该语音网络模型的输入，以该每个样本语音数据段的标注标签为该语音网络模型的训练目标，对该语音网络模块进行迭代训练。
在步骤S201和S202中，为了提高语音网络模型的预测准确度，可以对语音网络模型进行优化训练。具体的，电子设备可以采集训练样本集，该训练样本集包括样本音频数据的多个样本语音数据段，以及每个样本语音数据段的标注标签。其中，该标注标签包括样本语音数据段为边界点的概率，该标注标签可以是指人工对特征样本进行标注的。为了提高语音网络模型的适用范围，该样本音频数据可以是由处于不同区域的对象的语音数据段组成，或/和，该样本音频数据可以是由不同年龄的对象的语音数据段组成。进一步，以该多个样本语音数据段为该语音网络模型的输入，以该每个样本语音数据段的标注标签为该语音网络模型的训练目标；当该语音网络模型输出的语音数据段的预测标签与该语音数据段的标注标签相同，或者这两者相似时，表明语音网络模型的预测准确度较高，则结束对语音网络模型的迭代训练；当该语音网络模型输出的语音数据段的预测标签与该语音数据段的标注标签差异较大，表明该语音网络模型的预测准确度较低，则调整语音网络模型的网络参数，继续对该语音网络模型进行迭代训练。
在一个实施例中，步骤S202包括：将该多个样本语音数据段输入到该语音网络模型进行预测，得到每个样本语音数据段的预测标签；根据该每个样本语音数据段的预测标签及对应的样本语音数据段的预测标签确定该语音网络模型的预测误差；若该语音网络模型的预测误差大于预设误差值，则调整该语音网络模型的网络参数；若该语音网络模型的预测误差小于或等于该预设误差值，则结束对该语音网络模型的迭代训练。
电子设备可以通过调整语音网络模型的网络参数来优化语音网络模型，具体的，将该多个样本语音数据段输入到该语音网络模型进行预测，得到每个样本语音数据段的预测标签，根据该每个样本语音数据段的预测标签及对应的样本语音数据段的标注标签确定该语音网络模型的预测误差；若该语音网络模型的预测误差大于预设误差值，表明语音网络模型的预测准确度较低，则调整该语音网络模型的网络参数，并继续对语音网络模型进行迭代训练；若该语音网络模型的预测误差小于或等于该预设误差值，则表明语音网络模型的预测准确度较高，可以结束对该语音网络模型的迭代训练。例如，假设目标语音数据段包括T个语音数据段，可表示为X＝(X1，X2，...XT)，其中，每个语音数据段可以是由多个特征点组成，第i个语音数据段的标注标签为y i，第i个语音数据段的预测标签为f(X) i，则语音网络模型的预测误差L可以表示为公式(1)所示。
Figure PCTCN2019088976-appb-000001
在一个实施例中，接收将目标样本语音数据段的标注标签设置为第一类标注标签的指令，该第一类标注标签为样本语音数据段为边界点的概率大于预设概率 的标签，该目标样本语音数据段为样本音频数据的边界点所在的语音数据段；将与该目标样本语音数据段之间的时间间隔小于或等于预设时间间隔的样本语音数据段的标注标签设置为第一类标注标签；将与该目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段的标注标签设置为第二类标注标签，该第二类标注标签为样本语音数据段为边界点的概率小于或等于该预设概率的标签。
语音数据中包括边界点和非边界点，由于语音数据中边界点相对比较少，即只有极少的语音数据段为边界点的概率大于预设阈值，极多数的语音数据段为边界点的概率小于预设阈值；这两类数据的不平衡(相差太多)特征在对语音网络模型进行训练时导致很多问题，如导致语音网络模型的预测准确度较低。因此，可以将为真正边界点的附近增加正样本的数量，即增加概率大于预设概率的语音数据段。具体的，电子设备可以接收将目标样本语音数据段的标注标签设置为第一类标注标签的指令，该目标样本语音数据段为样本音频数据的边界点所在的语音数据段，即目标样本语音数据段为真正的边界点所在的语音数据段。为了解决不平衡问题，可以将与该目标样本语音数据段之间的时间间隔小于或等于预设时间间隔的样本语音数据段的标注标签设置为第一类标注标签，即将目标样本语音数据段附近的语音数据段都标注为边界点(即正样本)所在的语音数据段。将与该目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段的标注标签设置为第二类标注标签，即将与该目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段标注为非边界点所在的语音数据段。例如，预设时间间隔为50ms，电子设备可以将与目标样本数据段的时间间隔为50ms内的所有样本语音数据段标注为第一类标注标签。
S203、获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成。
S204、按照预设的处理规则对待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则。
S205、对该目标语音数据进行划分，得到多个语音数据段，将该多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，该预测标签包括语音数据段为边界点的概率。
S206、根据该每个语音数据段的预测标签确定该目标语音数据的边界点，以根据该边界点从该目标语音数据中分割出每个对象的语音数据段。
在一个实施例中，该每个语音数据段包括多个特征点，每相邻两个语音数据段包括相同的特征点，该每个语音数据段的预测标签为对应语音数据段内每个特征点的预测标签，步骤S206包括：根据该每个语音数据段的预测标签统计每个特征点为边界点的平均概率，将平均概率大于预设概率的所有特征点作为该目标语音数据的边界点。
由于相邻的每个语音数据段具有重叠性，因此，每相邻两个语音数据段包括相同的特征点，针对同一个特征点存在有两个概率的情况下，电子设备可以根据各个特征点的平均概率来确定语音数据的边界点。例如，假设目标语音数据包括第一语音数据段及第二语音数据段，第一语音数据段包括特征点x1，x2，x3，x4及x5，第一语音数据段的预测标签为0.2，则特征点x1，x2，x3，x4及x5的预测标签(概率)均为0.2；第二语音数据段包括特征点x4，x5，x6，x7及x8。第二语音数据段的预测标签为0.5，则特征点x4，x5，x6，x7及x8的预测标签均为0.5；根据预测标签确定每个特征点的平均标签(即平均概率)，即特 征点x1，x2，x3的平均预测标签均为0.2，特征点x4及x5的平均预测标签为0.35，特征点x6，x7及x8的平均预测标签需要根据第三语音数据段的预测标签进行计算。进一步，将平均概率大于预设概率的所有特征点作为该目标语音数据的边界点。
本申请实施例中，可通过语音网络模型自动获取目标语音数据的边界点，不需要人工操作，可节省大量人力，并可提高获取语音数据的边界点的准确度及效率；满足用户对语音数据处理的智能化、自动化需求。
请参见图3，是本申请实施例提供的一种语音数据处理装置的结构示意图，本申请实施例的该装置可以设置在上述提及的电子设备中。本实施例中，该装置包括：
获取单元301，用于获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成。
处理单元302，用于按照预设的处理规则对待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则。
预测单元303，用于对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率。
分割单元304，用于根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
可选的，处理单元302，用于按照预设步长对所述目标语音数据进行划分，得到多个语音数据段，所述多个语音数据段中每个语音数据段的长度大于所述预设步长。
可选的，训练单元305，用于获取训练样本集合，所述训练样本集合包括样本音频数据的多个样本语音数据段，及每个样本语音数据段的标注标签，所述标注标签包括样本语音数据段为边界点的概率；以所述多个样本语音数据段为所述语音网络模型的输入，以所述每个样本语音数据段的标注标签为所述语音网络模型的训练目标，对所述语音网络模块进行迭代训练。
可选的，训练单元305，用于将所述多个样本语音数据段输入到所述语音网络模型进行预测，得到每个样本语音数据段的预测标签；根据所述每个样本语音数据段的预测标签及对应的样本语音数据段的标注标签确定所述语音网络模型的预测误差；若所述语音网络模型的预测误差大于预设误差值，则调整所述语音网络模型的网络参数；若所述语音网络模型的预测误差小于或等于所述预设误差值，则结束对所述语音网络模型的迭代训练。
可选的，训练单元305，用于接收将目标样本语音数据段的标注标签设置为第一类标注标签的指令，所述第一类标注标签为样本语音数据段为边界点的概率大于预设概率的标签，所述目标样本语音数据段为样本音频数据的边界点所在的语音数据段；将与所述目标样本语音数据段之间的时间间隔小于或等于预设时间间隔的样本语音数据段的标注标签设置为第一类标注标签；将与所述目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段的标注标签设置为第二类标注标签，所述第二类标注标签为样本语音数据段为边界点的概率小于或等于所述预设概率的标签。
可选的，所述每个语音数据段包括多个特征点，每相邻两个语音数据段包括相同的特征点，所述每个语音数据段的预测标签为对应语音数据段内每个特征点的预测标签；分割单元304，用于根据所述每个语音数据段的预测标签统计每个特征点为边界点的平均概率；将平均概率大于预设概率的所有特征点作为所述目标语音数据的边界点。
可选的，所述预设的处理规则包括数据过滤规则，处理单元302，用于将所述待处理的语音数据进行划分，得到多个原始语音数据段；获取所述多个原始语音数据段中每个原始语音数据段的能量值；将所述多个原始语音数据段中能量值小于或等于预设能量值的原始语音数据段删除；将所述多个原始语音数据段中能量值大于所述预设能量值的原始语音数据段进行合并，得到所述目标语音数据。
本申请实施例中，可通过语音网络模型自动获取目标语音数据的边界点，不需要人工操作，可节省大量人力，并可提高获取语音数据的边界点的准确度及效率；满足用户对语音数据处理的智能化、自动化需求。
请参见图4，是本申请实施例提供的一种电子设备的结构示意图，如图所示的本实施例中的电子设备可以包括：一个或多个处理器401；一个或多个输入装置402，一个或多个输出装置403和存储器404。上述处理器401、输入装置402、输出装置403和存储器404通过总线404连接。
所处理器401可以是中央处理单元(Central Processing Unit，CPU)，该处理器还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
输入装置402可以包括触控板、指纹采传感器(用于采集用户的指纹信息和指纹的方向信息)、麦克风等，输出装置403可以包括显示器(LCD等)、扬声器等，输出装置403可以输出提示信息，提示信息可用于提示目标语音数据的边界点。
该存储器404可以包括只读存储器和随机存取存储器，并向处理器401提供指令和数据。存储器404的一部分还可以包括非易失性随机存取存储器，存储器404用于存储计算机程序，该计算机程序包括程序指令，处理器401用于执行存储器404存储的程序指令，以用于执行一种语音数据处理方法，即用于执行以下操作：
获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
按照预设的处理规则对所述待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
可选的，处理器401用于执行存储器404存储的程序指令，用于执行以下操作：
按照预设步长对所述目标语音数据进行划分，得到多个语音数据段，所述多个语音数据段中每个语音数据段的长度大于所述预设步长。
可选的，处理器401用于执行存储器404存储的程序指令，用于执行以下操作：
获取训练样本集合，所述训练样本集合包括样本音频数据的多个样本语音数据段，及每个样本语音数据段的标注标签，所述标注标签包括样本语音数据段为边界点的概率；
以所述多个样本语音数据段为所述语音网络模型的输入，以所述每个样本语音数据段的标注标签为所述语音网络模型的训练目标，对所述语音网络模块进行迭代训练。
可选的，处理器401用于执行存储器404存储的程序指令，用于执行以下操作：
将所述多个样本语音数据段输入到所述语音网络模型进行预测，得到每个样本语音数据段的预测标签；
根据所述每个样本语音数据段的预测标签及对应的样本语音数据段的标注标签确定所述语音网络模型的预测误差；
若所述语音网络模型的预测误差大于预设误差值，则调整所述语音网络模型的网络参数；
若所述语音网络模型的预测误差小于或等于所述预设误差值，则结束对所述语音网络模型的迭代训练。
可选的，处理器401用于执行存储器404存储的程序指令，用于执行以下操作：
接收将目标样本语音数据段的标注标签设置为第一类标注标签的指令，所述第一类标注标签为样本语音数据段为边界点的概率大于预设概率的标签，所述目标样本语音数据段为样本音频数据的边界点所在的语音数据段；
将与所述目标样本语音数据段之间的时间间隔小于或等于预设时间间隔的样本语音数据段的标注标签设置为第一类标注标签；
将与所述目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段的标注标签设置为第二类标注标签，所述第二类标注标签为样本语音数据段为边界点的概率小于或等于所述预设概率的标签。
可选的，处理器401用于执行存储器404存储的程序指令，用于执行以下操作：
根据所述每个语音数据段的预测标签统计每个特征点为边界点的平均概率；
将平均概率大于预设概率的所有特征点作为所述目标语音数据的边界点。
可选的，处理器401用于执行存储器404存储的程序指令，用于执行以下操作：
将所述待处理的语音数据进行划分，得到多个原始语音数据段；
获取所述多个原始语音数据段中每个原始语音数据段的能量值；
将所述多个原始语音数据段中能量值小于或等于预设能量值的原始语音数据段删除；
将所述多个原始语音数据段中能量值大于所述预设能量值的原始语音数据段进行合并，得到所述目标语音数据。
本申请实施例中所描述的处理器401、输入装置402、输出装置403可执行本申请实施例提供的语音数据处理方法的第一实施例和第二实施例中所描述的 实现方式，也可执行本申请实施例所描述的电子设备的实现方式，在此不再赘述。
本申请实施例中提供还了一种计算机可读存储介质，该计算机可读存储介质存储有计算机程序，该计算机程序包括程序指令，该程序指令被处理器执行时实现本申请的图1及图2实施例中所示的语音数据处理方法。
该计算机可读存储介质可以是前述任一实施例该的医疗管理设备的内部存储单元，例如控制设备的硬盘或内存。该计算机可读存储介质也可以是该控制设备的外部存储设备，例如该控制设备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，该计算机可读存储介质还可以既包括该控制设备的内部存储单元也包括外部存储设备。该计算机可读存储介质用于存储该计算机程序以及该控制设备所需的其他程序和数据。该计算机可读存储介质还可以用于暂时地存储已经输出或者将要输出的数据。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，上述描述的控制设备和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的控制设备和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例是示意性的，例如，该单元的划分，可以为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。
以上该，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到各种等效的修改或替换，这些修改或替换都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以权利要求的保护范围为准。

Claims (20)

  1. 一种语音数据处理方法，其特征在于，包括：
    获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
    按照预设的处理规则对所述待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
    对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
    根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
  2. 根据权利要求1所述的方法，其特征在于，所述对所述目标语音数据进行划分，得到多个语音数据段，包括：
    按照预设步长对所述目标语音数据进行划分，得到多个语音数据段，所述多个语音数据段中每个语音数据段的长度大于所述预设步长。
  3. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    获取训练样本集合，所述训练样本集合包括样本音频数据的多个样本语音数据段，及每个样本语音数据段的标注标签，所述标注标签包括样本语音数据段为边界点的概率；
    以所述多个样本语音数据段为所述语音网络模型的输入，以所述每个样本语音数据段的标注标签为所述语音网络模型的训练目标，对所述语音网络模块进行迭代训练。
  4. 根据权利要求3所述的方法，其特征在于，所述以所述多个样本语音数据段为所述语音网络模型的输入，以所述每个样本语音数据段的标注标签为所述语音网络模型的训练目标，对所述语音网络模块进行迭代训练，包括：
    将所述多个样本语音数据段输入到所述语音网络模型进行预测，得到每个样本语音数据段的预测标签；
    根据所述每个样本语音数据段的预测标签及对应的样本语音数据段的标注标签确定所述语音网络模型的预测误差；
    若所述语音网络模型的预测误差大于预设误差值，则调整所述语音网络模型的网络参数；
    若所述语音网络模型的预测误差小于或等于所述预设误差值，则结束对所述语音网络模型的迭代训练。
  5. 根据权利要求3或4所述的方法，其特征在于，所述方法还包括：
    接收将目标样本语音数据段的标注标签设置为第一类标注标签的指令，所述第一类标注标签为样本语音数据段为边界点的概率大于预设概率的标签，所述目标样本语音数据段为样本音频数据的边界点所在的语音数据段；
    将与所述目标样本语音数据段之间的时间间隔小于或等于预设时间间隔的样本语音数据段的标注标签设置为第一类标注标签；
    将与所述目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段的标注标签设置为第二类标注标签，所述第二类标注标签为样本语音数据段为边界点的概率小于或等于所述预设概率的标签。
  6. 根据权利要求1所述的方法，其特征在于，所述每个语音数据段包括多个特征点，每相邻两个语音数据段包括相同的特征点，所述每个语音数据段的预 测标签为对应语音数据段内每个特征点的预测标签；
    所述根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，包括：
    根据所述每个语音数据段的预测标签统计每个特征点为边界点的平均概率；
    将平均概率大于预设概率的所有特征点作为所述目标语音数据的边界点。
  7. 根据权利要求1所述的方法，其特征在于，所述预设的处理规则包括数据过滤规则，所述按照预设的处理规则对所述待处理的语音数据进行处理，得到目标语音数据，包括：
    将所述待处理的语音数据进行划分，得到多个原始语音数据段；
    获取所述多个原始语音数据段中每个原始语音数据段的能量值；
    将所述多个原始语音数据段中能量值小于或等于预设能量值的原始语音数据段删除；
    将所述多个原始语音数据段中能量值大于所述预设能量值的原始语音数据段进行合并，得到所述目标语音数据。
  8. 一种语音数据处理装置，其特征在于，包括：
    获取单元，用于获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
    处理单元，用于按照预设的处理规则对待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
    预测单元，用于对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
    分割单元，用于根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
  9. 根据权利要求8所述的装置，其特征在于，所述预测单元，具体用于按照预设步长对所述目标语音数据进行划分，得到多个语音数据段，所述多个语音数据段中每个语音数据段的长度大于所述预设步长。
  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：
    训练单元，用于获取训练样本集合，所述训练样本集合包括样本音频数据的多个样本语音数据段，及每个样本语音数据段的标注标签，所述标注标签包括样本语音数据段为边界点的概率；以所述多个样本语音数据段为所述语音网络模型的输入，以所述每个样本语音数据段的标注标签为所述语音网络模型的训练目标，对所述语音网络模块进行迭代训练。
  11. 根据权利要求10所述的装置，其特征在于，所述训练单元，具体用于将所述多个样本语音数据段输入到所述语音网络模型进行预测，得到每个样本语音数据段的预测标签；根据所述每个样本语音数据段的预测标签及对应的样本语音数据段的标注标签确定所述语音网络模型的预测误差；若所述语音网络模型的预测误差大于预设误差值，则调整所述语音网络模型的网络参数；若所述语音网络模型的预测误差小于或等于所述预设误差值，则结束对所述语音网络模型的迭代训练。
  12. 根据权利要求10或11所述的装置，其特征在于，
    所述训练单元，还用于接收将目标样本语音数据段的标注标签设置为第一 类标注标签的指令，所述第一类标注标签为样本语音数据段为边界点的概率大于预设概率的标签，所述目标样本语音数据段为样本音频数据的边界点所在的语音数据段；将与所述目标样本语音数据段之间的时间间隔小于或等于预设时间间隔的样本语音数据段的标注标签设置为第一类标注标签；将与所述目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段的标注标签设置为第二类标注标签，所述第二类标注标签为样本语音数据段为边界点的概率小于或等于所述预设概率的标签。
  13. 根据权利要求8所述的装置，其特征在于，所述每个语音数据段包括多个特征点，每相邻两个语音数据段包括相同的特征点，所述每个语音数据段的预测标签为对应语音数据段内每个特征点的预测标签；
    所述预测单元，具体用于根据所述每个语音数据段的预测标签统计每个特征点为边界点的平均概率；将平均概率大于预设概率的所有特征点作为所述目标语音数据的边界点。
  14. 根据权利要求8所述的装置，其特征在于，所述预设的处理规则包括数据过滤规则，所述处理单元，具体用于将所述待处理的语音数据进行划分，得到多个原始语音数据段；获取所述多个原始语音数据段中每个原始语音数据段的能量值；将所述多个原始语音数据段中能量值小于或等于预设能量值的原始语音数据段删除；将所述多个原始语音数据段中能量值大于所述预设能量值的原始语音数据段进行合并，得到所述目标语音数据。
  15. 一种电子设备，其特征在于，包括：
    处理器，适于实现一条或一条以上指令；以及，
    计算机可读存储介质，所述计算机可读存储介质存储有一条或一条以上指令，所述一条或一条以上指令适于由所述处理器加载并执行如下步骤：获取待处理的语音数据，所述待处理的语音数据由多个对象的语音数据段组成；
    按照预设的处理规则对所述待处理的语音数据进行处理，得到目标语音数据，所述预设的处理规则包括数据过滤规则和/或数据格式处理规则；
    对所述目标语音数据进行划分，得到多个语音数据段；将所述多个语音数据段输入到语音网络模型中进行预测，得到每个语音数据段的预测标签，所述预测标签包括语音数据段为边界点的概率；
    根据所述每个语音数据段的预测标签确定所述目标语音数据的边界点，以根据所述边界点从所述目标语音数据中分割出每个对象的语音数据段。
  16. 根据权利要求15所述的电子设备，其特征在于，所述处理器，具体用于按照预设步长对所述目标语音数据进行划分，得到多个语音数据段，所述多个语音数据段中每个语音数据段的长度大于所述预设步长。
  17. 根据权利要求15所述的电子设备，其特征在于，所述处理器，具体用于获取训练样本集合，所述训练样本集合包括样本音频数据的多个样本语音数据段，及每个样本语音数据段的标注标签，所述标注标签包括样本语音数据段为边界点的概率；以所述多个样本语音数据段为所述语音网络模型的输入，以所述每个样本语音数据段的标注标签为所述语音网络模型的训练目标，对所述语音网络模块进行迭代训练。
  18. 根据权利要求17所述的电子设备，其特征在于，所述处理器，具体用于将所述多个样本语音数据段输入到所述语音网络模型进行预测，得到每个样本语音数据段的预测标签；根据所述每个样本语音数据段的预测标签及对应的样本语音数据段的标注标签确定所述语音网络模型的预测误差；若所述语音网络模型 的预测误差大于预设误差值，则调整所述语音网络模型的网络参数；若所述语音网络模型的预测误差小于或等于所述预设误差值，则结束对所述语音网络模型的迭代训练。
  19. 根据权利要求17或18所述的电子设备，其特征在于，所述处理器，还用于接收将目标样本语音数据段的标注标签设置为第一类标注标签的指令，所述第一类标注标签为样本语音数据段为边界点的概率大于预设概率的标签，所述目标样本语音数据段为样本音频数据的边界点所在的语音数据段；将与所述目标样本语音数据段之间的时间间隔小于或等于预设时间间隔的样本语音数据段的标注标签设置为第一类标注标签；将与所述目标样本语音数据段之间的时间间隔大于预设时间间隔的样本语音数据段的标注标签设置为第二类标注标签，所述第二类标注标签为样本语音数据段为边界点的概率小于或等于所述预设概率的标签。
  20. 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存储有一条或一条以上指令，所述一条或一条以上指令适于由处理器加载并执行如权利要求1-7任一项所述的语音数据处理方法。
PCT/CN2019/088976 2019-01-04 2019-05-29 语音数据处理方法、装置、设备及存储介质 Ceased WO2020140374A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910018423.9A CN109616097B (zh) 2019-01-04 2019-01-04 语音数据处理方法、装置、设备及存储介质
CN201910018423.9 2019-01-04

Publications (1)

Publication Number Publication Date
WO2020140374A1 true WO2020140374A1 (zh) 2020-07-09

Family

ID=66018221

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/088976 Ceased WO2020140374A1 (zh) 2019-01-04 2019-05-29 语音数据处理方法、装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN109616097B (zh)
WO (1) WO2020140374A1 (zh)

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109616097B (zh) * 2019-01-04 2024-05-10 平安科技（深圳）有限公司 语音数据处理方法、装置、设备及存储介质
CN111145765B (zh) * 2019-12-31 2022-04-15 思必驰科技股份有限公司 一种音频处理方法、装置、电子设备及存储介质
CN111312223B (zh) * 2020-02-20 2023-06-30 北京声智科技有限公司 语音分割模型的训练方法、装置和电子设备
CN111312224B (zh) * 2020-02-20 2023-04-21 北京声智科技有限公司 语音分割模型的训练方法、装置和电子设备
CN111785302B (zh) * 2020-06-23 2024-08-20 北京声智科技有限公司 说话人分离方法、装置及电子设备
CN113035234B (zh) * 2021-03-10 2024-02-09 湖南快乐阳光互动娱乐传媒有限公司 音频数据处理方法及相关装置
CN113593528B (zh) * 2021-06-30 2022-05-17 北京百度网讯科技有限公司 语音切分模型的训练方法、装置、电子设备及存储介质
CN114187924B (zh) * 2021-12-08 2025-10-28 北京百度网讯科技有限公司 数据处理方法、装置、电子设备及介质
CN116052714A (zh) * 2022-11-30 2023-05-02 华为技术有限公司 一种数据处理方法及其装置
CN115835172B (zh) * 2022-12-01 2026-01-20 深圳市中科蓝讯科技股份有限公司 一种蓝牙数据传输方法、主设备及从设备
CN116432111B (zh) * 2023-04-18 2025-12-23 中国工商银行股份有限公司 资源数据处理方法、装置和计算机设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160111112A1 (en) * 2014-10-17 2016-04-21 Fujitsu Limited Speaker change detection device and speaker change detection method
CN106782507A (zh) * 2016-12-19 2017-05-31 平安科技（深圳）有限公司 语音分割的方法及装置
US20180039888A1 (en) * 2016-08-08 2018-02-08 Interactive Intelligence Group, Inc. System and method for speaker change detection
CN109616097A (zh) * 2019-01-04 2019-04-12 平安科技（深圳）有限公司 语音数据处理方法、装置、设备及存储介质

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4175093B2 (ja) * 2002-11-06 2008-11-05 日本電信電話株式会社 トピック境界決定方法及び装置及びトピック境界決定プログラム
US20080189109A1 (en) * 2007-02-05 2008-08-07 Microsoft Corporation Segmentation posterior based boundary point determination
CN104409080B (zh) * 2014-12-15 2018-09-18 北京国双科技有限公司 语音端点检测方法和装置
CN107342077A (zh) * 2017-05-27 2017-11-10 国家计算机网络与信息安全管理中心 一种基于因子分析的说话人分段聚类方法及系统
CN108257592A (zh) * 2018-01-11 2018-07-06 广州势必可赢网络科技有限公司 一种基于长短期记忆模型的人声分割方法及系统
CN108417201B (zh) * 2018-01-19 2020-11-06 苏州思必驰信息科技有限公司 单信道多说话人身份识别方法及系统
CN108766418B (zh) * 2018-05-24 2020-01-14 百度在线网络技术（北京）有限公司 语音端点识别方法、装置及设备
CN108986844B (zh) * 2018-08-06 2020-08-28 东北大学 一种基于说话人语音特征的语音端点检测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160111112A1 (en) * 2014-10-17 2016-04-21 Fujitsu Limited Speaker change detection device and speaker change detection method
US20180039888A1 (en) * 2016-08-08 2018-02-08 Interactive Intelligence Group, Inc. System and method for speaker change detection
CN106782507A (zh) * 2016-12-19 2017-05-31 平安科技（深圳）有限公司 语音分割的方法及装置
CN109616097A (zh) * 2019-01-04 2019-04-12 平安科技（深圳）有限公司 语音数据处理方法、装置、设备及存储介质

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
RUIQING YIN ET AL: "Speaker Change Detection in Broadcast TV Using Bidirectional Long Short-Term Memory Networks", INTERSPEECH 2017, 24 August 2017 (2017-08-24), Stockholm, Sweden, pages 3827 - 3831, XP055716415, DOI: 10.21437/Interspeech.2017-65 *
ZHENHAO GE ET AL: "Speaker Change Detection Using Features through A Neural Network Speaker Classifier", INTELLIGENT SYSTEMS CONFERENCE 2017, 8 September 2017 (2017-09-08), pages 1 - 6, XP080955924, DOI: 10.1109/IntelliSys.2017.8324268 *

Also Published As

Publication number Publication date
CN109616097A (zh) 2019-04-12
CN109616097B (zh) 2024-05-10

Similar Documents

Publication Publication Date Title
WO2020140374A1 (zh) 语音数据处理方法、装置、设备及存储介质
CN111523640B (zh) 神经网络模型的训练方法和装置
US11887623B2 (en) End-to-end speech diarization via iterative speaker embedding
CN110598620B (zh) 基于深度神经网络模型的推荐方法和装置
WO2019196278A1 (zh) 天气数据获取方法及装置、计算机装置及可读存储介质
CN108172213A (zh) 娇喘音频识别方法、装置、设备及计算机可读介质
CN109087667B (zh) 语音流利度识别方法、装置、计算机设备及可读存储介质
CN111160959B (zh) 一种用户点击转化预估方法及装置
WO2022078002A1 (zh) 一种图像处理方法、装置、设备及可读存储介质
CN116091869A (zh) 模型训练方法、图像标注方法、装置、电子设备及介质
CN118537666A (zh) 图像检测方法、装置、电子设备及存储介质
CN118823630A (zh) 一种视频理解方法、系统、电子设备和存储介质
CN112989869A (zh) 人脸质量检测模型的优化方法、装置、设备及存储介质
CN113762403B (zh) 图像处理模型量化方法、装置、电子设备及存储介质
CN116913325A (zh) 一种噪声事件检测方法和装置
WO2022183363A1 (zh) 一种模型的训练方法、装置、终端设备及存储介质
CN115050350A (zh) 标注检查方法及相关装置、电子设备、存储介质
CN111785259A (zh) 信息处理方法、装置及电子设备
US20220130405A1 (en) Low Complexity Voice Activity Detection Algorithm
CN115879002B (zh) 一种训练样本生成方法、模型训练方法及装置
CN119629636A (zh) 垃圾电话的识别方法、装置、计算机设备及存储介质
CN107071553A (zh) 一种修改视频语音的方法、装置和计算机可读存储介质
CN114049899B (zh) 一种声音识别方法、装置、电子设备及存储介质
CN114881994B (zh) 产品缺陷检测方法、装置以及存储介质
CN112966110B (zh) 文本类别识别方法及相关设备

Legal Events

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

Ref document number: 19907856

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: 19907856

Country of ref document: EP

Kind code of ref document: A1