A method and system for generating precision-standardized data based on unsupervised data
By employing unsupervised data generation methods and utilizing VAD phrase filtering, speaker classification, and forced alignment techniques, the efficiency and cost issues of acquiring training data for speech recognition were resolved, achieving efficient and low-cost acquisition of accurate labeled data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI SHENFEI TECHNOLOGY CO LTD
- Filing Date
- 2025-07-08
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies face challenges in acquiring training data for speech recognition due to increasing data volume requirements and the need for precise data acquisition, resulting in high costs and low efficiency.
An unsupervised data generation method is adopted, which performs quality screening and speaker classification by acquiring VAD phrases, applies random noise, performs speech recognition and forced alignment, and uses signal-to-noise ratio estimation and confidence screening to obtain high-confidence audio data.
It improves the production efficiency of precise speech recognition data, reduces acquisition costs, and enables the acquisition of a large amount of high-quality training data in a short period of time.
Smart Images
Figure CN120998190B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of text processing and speech recognition technology, specifically a method and system for generating precise labeled data based on unsupervised data. Background Technology
[0002] As artificial intelligence products become increasingly integrated into daily life, speech recognition, as a fundamental means of natural human-computer interaction, has found wide application in scenarios such as smart homes, smart cars, and intelligent customer service. As a prerequisite for natural human-computer interaction, speech recognition is inevitably a development trend in artificial intelligence; only when the accuracy of speech recognition approaches perfection can natural human-computer interaction continue.
[0003] With the rapid development of artificial intelligence technology, especially the leaps and bounds in large model technology, the demand for training data in speech recognition technology has increased significantly, from tens of thousands of hours of training data for commercial models to hundreds of thousands of hours. This has brought challenges to the methods of acquiring accurate speech data. Summary of the Invention
[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a method and system for generating precise calibration data based on unsupervised data, thereby enabling the acquisition of a large amount of acoustic model training data in a short period of time.
[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a method for generating precise standardization data based on unsupervised data, the method comprising the following steps:
[0006] The process involves acquiring VAD phrases, performing quality filtering on the VAD phrases to obtain high-quality VAD phrases, classifying and filtering the high-quality VAD phrases by speaker to obtain filtered single-speaker VAD phrases, and applying random noise to the single-speaker VAD phrases to obtain noisy VAD phrases.
[0007] Speech recognition tools are used to identify single-person spoken VAD phrases and noisy VAD phrases to obtain the corresponding text of VAD phrases and noisy VAD phrases.
[0008] Force alignment is performed on the texts of single-person speaking VAD phrases and their corresponding texts, as well as the texts of noisy VAD phrases and their corresponding texts, and the confidence scores of the alignment are obtained to obtain the filtering criteria confidence_1 and confidence_2.
[0009] The audio is received and filtered based on the filtering criteria confidence_1 and confidence_2 to obtain audio with high confidence.
[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: performing quality screening on the VAD phrases to obtain high-audio-quality VAD phrases, including:
[0011] By using a signal-to-noise ratio estimation method, sentences with poor audio quality are removed, resulting in high-quality VAD short sentences.
[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: performing speaker classification and filtering on the high-audio-quality VAD phrases, including:
[0013] Each sentence within a high-quality VAD phrase is segmented and the speaker is categorized. For sentences containing multiple speakers, the phrase is truncated again at the speaker switching point for further filtering to ensure that each sentence has only one speaker, resulting in multiple single-speaker VAD phrases.
[0014] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of recognizing single-person spoken VAD phrases and noisy VAD phrases using a speech recognition tool includes:
[0015] words1 = F(s1)
[0016] words2 = F(s2)
[0017] F represents speech recognition technology; s1 represents a short VAD sentence spoken by a single person; s2 represents a short sentence after adding noise to s1; words1 is the text sentence converted from s1 using speech recognition technology; words2 is the text sentence converted from s2 using speech recognition technology.
[0018] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the use of force-alignment technology for forced alignment and obtaining the confidence level of the alignment, including:
[0019] confidence_1 = P(s1|words 1)
[0020] confidence_2 = P(s1|words2)
[0021] confidence_1 is the probability of a given speech signal s1 in the context of the text sentence words1, and confidence_2 is the probability of a given speech signal s1 in the context of the text sentence words2.
[0022] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of filtering audio based on the filtering criteria confidence_1 and confidence_2 includes: when confidence_1>=confidence_2 and confidence_1>thresh, words1 is taken as the finely labeled text; when confidence_1<=confidence_2 and confidence_2>thresh, words2 is taken as the finely labeled text; in other cases, it indicates that the confidence of the recognition result text is not high and there is a certain risk, so the sentence is discarded.
[0023] Secondly, in order to achieve the above objectives, this invention discloses a system for generating precise calibration data based on unsupervised data, comprising:
[0024] The sentence filtering module is used to obtain VAD sentences, perform quality filtering on the VAD sentences to obtain high-audio quality VAD sentences; perform speaker classification filtering on the high-audio quality VAD sentences to obtain filtered single-speaker VAD sentences; and apply random noise to the single-speaker VAD sentences to obtain noisy VAD sentences.
[0025] The speech recognition module is used to recognize VAD phrases spoken by a single person and noisy VAD phrases using a speech recognition tool, and to obtain the corresponding text of the VAD phrases and the corresponding text of the noisy VAD phrases.
[0026] The text alignment module is used to force-align the text of single-person speaking VAD phrases and their corresponding texts, as well as the text of noisy VAD phrases and their corresponding texts, using force-alignment technology, and obtain the alignment confidence to obtain the filtering criteria confidence_1 and confidence_2.
[0027] The audio filtering module receives audio and filters it based on the filtering criteria confidence_1 and confidence_2 to obtain audio with high confidence.
[0028] In another aspect of the present invention, in order to achieve the above-mentioned objective, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs a method for generating precise data based on unsupervised data as described above.
[0029] In another aspect of the present invention, in order to achieve the above-mentioned objective, a computer-readable storage medium is disclosed, wherein a computer program is stored in the computer program, and when the computer program is loaded and executed by a processor, a method for generating precision data based on unsupervised data as described above is employed.
[0030] The beneficial effects of this invention are:
[0031] In the process of producing a large amount of precise labeled speech recognition data, this invention uses multiple screening and verification methods to obtain reliable labels from unsupervised data. Compared with traditional methods for obtaining precise labeled data, this improves data production efficiency and reduces costs. Attached Figure Description
[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0034] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0036] Example 1:
[0037] like Figure 1 As shown, a method for generating precise standardization data based on unsupervised data includes the following steps:
[0038] S101: Obtain VAD phrases, perform quality screening on VAD phrases to obtain high-quality VAD phrases; perform speaker classification screening on high-quality VAD phrases to obtain screened single-speaker VAD phrases; apply random noise to single-speaker VAD phrases to obtain noisy VAD phrases.
[0039] The process of quality screening for VAD phrases to obtain high-quality VAD phrases includes:
[0040] By using a signal-to-noise ratio estimation method, sentences with poor audio quality are removed, resulting in high-quality VAD short sentences.
[0041] The high-quality VAD phrases are used for speaker classification and filtering, including:
[0042] Each sentence within a high-quality VAD phrase is segmented and the speaker is categorized. For sentences containing multiple speakers, the phrase is truncated again at the speaker switching point for further filtering to ensure that each sentence has only one speaker, resulting in multiple single-speaker VAD phrases.
[0043] S102: Use a speech recognition tool to recognize single-person spoken VAD phrases and noisy VAD phrases to obtain the corresponding text of VAD phrases and the corresponding text of noisy VAD phrases;
[0044] The process of recognizing single-person spoken VAD phrases and noisy VAD phrases using a speech recognition tool includes:
[0045] words1 = F(s1)
[0046] words2 = F(s2)
[0047] F represents a speech recognition technology that can convert speech signals into text, generally referring to traditional speech recognition methods or end-to-end speech recognition methods;
[0048] s1 represents a vad phrase spoken by a single person;
[0049] s2 represents the short sentence after adding noise to s1;
[0050] words1 is a text sentence converted from s1 using speech recognition technology;
[0051] words2 is a text sentence converted from s2 using speech recognition technology.
[0052] S103: Force alignment is performed on the text of single-person speaking VAD phrases and their corresponding texts, as well as the text of noisy VAD phrases and their corresponding texts, using force-alignment technology. The confidence of the alignment is obtained to obtain the filtering criteria confidence_1 and confidence_2.
[0053] The process of using force-alignment technology to perform forced alignment and obtaining the confidence level of the alignment includes:
[0054] confidence_1 = P(s1|words 1)
[0055] confidence_2 = P(s1|words2)
[0056] confidence_1 is the probability of a given speech signal s1 in the context of a text sentence words1, and is usually calculated using the Viterbi method.
[0057] confidence_2 is the probability of a given speech signal s1 in the context of the text sentence words2.
[0058] S104: Receive audio, filter the audio based on the filtering criteria confidence_1 and confidence_2, and obtain audio with high confidence.
[0059] The process of filtering audio based on confidence_1 and confidence_2 includes:
[0060] When confidence_1 >= confidence_2 and confidence_1 > thresh, words1 is taken as the finely labeled text.
[0061] When confidence_1 <= confidence_2 and confidence_2 > thresh, words2 is taken as the precise labeled text; otherwise, it indicates that the confidence of the recognized text is not high and there is a certain risk, so the sentence is discarded.
[0062] Example 2: To achieve the above objective, such as Figure 2 As shown, based on Embodiment 1, this invention discloses a system for generating precision-calibrated data from unsupervised data, comprising:
[0063] The sentence filtering module 11 is used to obtain VAD short sentences, perform quality filtering on VAD short sentences to obtain high-audio quality VAD short sentences; perform speaker classification filtering on high-audio quality VAD short sentences to obtain filtered single-speaker VAD short sentences; and apply random noise to single-speaker VAD short sentences to obtain noisy VAD short sentences.
[0064] The speech recognition module 12 is used to recognize single-person spoken VAD phrases and noisy VAD phrases using a speech recognition tool, and obtain the corresponding text of the VAD phrases and the corresponding text of the noisy VAD phrases.
[0065] The text alignment module 13 is used to force-align the text of single-person speaking VAD phrases and corresponding VAD phrases, as well as the text of noisy VAD phrases and corresponding noisy VAD phrases, using force-alignment technology, and obtain the alignment confidence to obtain the filtering criteria confidence_1 and confidence_2.
[0066] The audio filtering module 14 is used to receive audio and filter the audio based on the filtering criteria confidence_1 and confidence_2 to obtain audio with high confidence.
[0067] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.
[0068] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
[0069] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0070] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.
Claims
1. A method for generating precise standardization data based on unsupervised data, characterized in that, The method includes the following steps: The process involves acquiring VAD phrases, performing quality filtering on the VAD phrases to obtain high-quality VAD phrases, classifying and filtering the high-quality VAD phrases by speaker to obtain filtered single-speaker VAD phrases, and applying random noise to the single-speaker VAD phrases to obtain noisy VAD phrases. Speech recognition tools are used to identify single-person spoken VAD phrases and noisy VAD phrases to obtain the corresponding text of VAD phrases and noisy VAD phrases. Force alignment is performed on the texts of single-person speaking VAD phrases and their corresponding texts, as well as the texts of noisy VAD phrases and their corresponding texts, and the confidence scores of the alignment are obtained to obtain the filtering criteria confidence_1 and confidence_2. The process of using force-alignment technology to perform forced alignment and obtaining the confidence level of the alignment includes: confidence_1 is the probability of a given speech signal s1 in the context of text sentence words1, and confidence_2 is the probability of a given speech signal s1 in the context of text sentence words2. The audio is received and filtered based on the filtering criteria confidence_1 and confidence_2 to obtain audio with high confidence. The process of filtering audio based on confidence_1 and confidence_2 includes: When confidence_1>=confidence_2 and confidence_1>thresh, words1 is taken as the finely labeled text; when confidence_1<=confidence_2 and confidence_2>thresh, words2 is taken as the finely labeled text; otherwise, it indicates that the confidence of the recognized text is not high and there is a risk, so the sentence is discarded.
2. The method for generating precise standard data based on unsupervised data according to claim 1, characterized in that, The process of quality screening for VAD phrases to obtain high-quality VAD phrases includes: By using a signal-to-noise ratio estimation method, sentences with poor audio quality are removed, resulting in high-quality VAD short sentences.
3. The method for generating precise standardization data based on unsupervised data according to claim 1, characterized in that, The high-quality VAD phrases are used for speaker classification and filtering, including: Each sentence within a high-quality VAD phrase is segmented and the speaker is categorized. For sentences containing multiple speakers, the phrase is truncated again at the speaker switching point for further filtering to ensure that each sentence has only one speaker, resulting in multiple single-speaker VAD phrases.
4. The method for generating precise standard data based on unsupervised data according to claim 1, characterized in that, The process of recognizing single-person spoken VAD phrases and noisy VAD phrases using a speech recognition tool includes: F represents speech recognition technology; s1 represents a single-person spoken VAD sentence; s2 represents a sentence after adding noise to s1, which is a noisy VAD sentence; words1 is the text sentence converted from s1 using speech recognition technology; words2 is the text sentence converted from s2 using speech recognition technology.
5. A system for generating precise standardization data based on unsupervised data, characterized in that, include: The sentence filtering module is used to obtain VAD sentences, perform quality filtering on the VAD sentences to obtain high-audio quality VAD sentences; perform speaker classification filtering on the high-audio quality VAD sentences to obtain filtered single-speaker VAD sentences; and apply random noise to the single-speaker VAD sentences to obtain noisy VAD sentences. The speech recognition module is used to recognize VAD phrases spoken by a single person and noisy VAD phrases using a speech recognition tool, and to obtain the corresponding text of the VAD phrases and the corresponding text of the noisy VAD phrases. The text alignment module is used to force-align the text of single-person speaking VAD phrases and their corresponding texts, as well as the text of noisy VAD phrases and their corresponding texts, using force-alignment technology, and obtain the alignment confidence to obtain the filtering criteria confidence_1 and confidence_2. Force alignment is performed using the force-alignment technique, and the confidence level of the alignment is obtained, including: confidence_1 is the probability of a given speech signal s1 in the context of text sentence words1, and confidence_2 is the probability of a given speech signal s1 in the context of text sentence words2. The audio filtering module receives audio and filters it based on the filtering criteria confidence_1 and confidence_2 to obtain audio with high confidence. The process of filtering audio based on confidence_1 and confidence_2 includes: When confidence_1>=confidence_2 and confidence_1>thresh, words1 is taken as the finely labeled text; when confidence_1<=confidence_2 and confidence_2>thresh, words2 is taken as the finely labeled text; otherwise, it indicates that the confidence of the recognized text is not high and there is a risk, so the sentence is discarded.
6. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs a method for generating precise standard data based on unsupervised data, as described in any one of claims 1 to 4.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs a method for generating precise standard data based on unsupervised data, as described in any one of claims 1 to 4.