# US 2019/0318723 A1 · Method for training voice data set, computer device, and computer-readable storage medium

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0318723 A1
- Document ID: US20190318723A1
- Status: Published application
- Kind: A1
- Application number: 16/436,479
- Filed: 2019-06-10
- Published: 2019-10-17
- Earliest claimed: 2017-03-10
- Assignees: Tencent Technology (Shenzhen) Company Limited
- Applicants: Tencent Technology (Shenzhen) Company Limited
- Inventors: Tao Sun; Yueteng Kang; Xiaoming Zhang; Li Zhang
- Web page: https://patentgrep.com/patent/US20190318723A1

## CPC

- G10L 15/063: {Training}
- G10L 15/02: Feature extraction for speech recognition; Selection of recognition unit
- G10L 15/04: Segmentation; Word boundary detection
- G10L 15/142: {Hidden Markov Models [HMMs]}
- G10L 15/144: {Training of HMMs}
- G10L 15/22: Procedures used during a speech recognition process, e.g. man-machine dialogue
- G10L 15/16: using artificial neural networks
- G10L 2015/025: {Phonemes, fenemes or fenones being the recognition units}
- G10L 2015/0635: {updating or merging of old and new templates; Mean values; Weighting}
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING

## Abstract

A method for training a voice data set is provided. A first test set of data selected from a first voice data set, and a first voice model parameter obtained by performing first voice model training based on a first voice data set, are obtained. Data from a second voice data set is randomly selected to generate a second test set. Further, second voice model training is performed based on the second voice data set and the first voice model parameter when the second test set and the first test set satisfy a similarity condition.

## Claims

### Claim 1

A training method, the method comprising:
obtaining, by processing circuitry of a computer device, a first test set of data selected from a first voice data set, and a first voice model parameter obtained by performing first voice model training based on a first voice data set;
randomly selecting, by the processing circuitry, data from a second voice data set to generate a second test set; and
performing, by the processing circuitry, second voice model training based on the second voice data set and the first voice model parameter when the second test set and the first test set satisfy a similarity condition.

### Claim 2 (depends on 1)

The method according to claim 1, wherein
the step of obtaining the first voice model parameter includes:
separately selecting data from the first voice data set to generate a training set and the first test set;
performing the first voice model training based on the training set to obtain a preset quantity of first voice models;
testing the first test set with the preset quantity of first voice models separately to obtain at least one first voice model having a word error rate within a preset range; and
the first voice model parameter is a parameter of one of the at least one first voice model having the word error rate within the preset range.

### Claim 3 (depends on 2)

The method according to claim 2, wherein the performing the first voice model training comprises:
randomly selecting a first preset proportion of data or a first fixed quantity of pieces of the data from the training set each time to perform the first voice model training, and repeating the first voice model training to obtain the preset quantity of first voice models.

### Claim 4 (depends on 2)

The method according to claim 2, wherein the testing the first test set comprises:
testing the first test set with the preset quantity of first voice models separately to obtain a word error rate of each of the first voice models; and
determining the at least one first voice model having the word error rate within the preset range according to the word error rate of each of the first voice models.

### Claim 5 (depends on 2)

The method according to claim 2, wherein the performing second voice model training comprises:
performing the second voice model training based on the second voice data set and the first voice model parameter of the first voice model having the word error rate being a minimum word error rate within the preset range.

### Claim 6 (depends on 5)

The method according to claim 5, further comprising:
testing, by the processing circuitry, the second test set with the first voice model corresponding to the word error rate being the minimum word error rate within the preset range to obtain a word error rate corresponding to the second test set, and
when a difference between the word error rate corresponding to the second test set and the minimum word error rate within the preset range is less than or equal to a tolerance threshold, determining, by the processing circuitry, that the second test set and the first test set satisfy the similarity condition.

### Claim 7 (depends on 1)

The method according to claim 1, wherein the randomly selecting data from the second voice data set comprises:
obtaining a ratio of a quantity of pieces of the data in the first test set to a quantity of pieces of data in the first voice data set, and randomly selecting random data from the second voice data set according to the ratio to generate the second test set.

### Claim 8 (depends on 1)

The method according to claim 1, further comprising:
separately selecting, by the processing circuitry, data from the first voice data set to generate a training set and the first test set;
performing, by the processing circuitry, the first voice model training based on the training set to obtain a preset quantity of first voice models;
testing, by the processing circuitry, the first test set with the preset quantity of first voice models separately to obtain a first voice model of the preset quantity of first voice models having a minimum word error rate; and
updating, by the processing circuitry, the first voice model parameter based on a parameter of the first voice model having the minimum word error rate.

### Claim 9 (depends on 8)

The method according to claim 8, further comprising:
testing, by the processing circuitry, the second test set with the first voice model having the minimum word error rate to obtain a word error rate corresponding to the second test set; and
when a difference between the word error rate corresponding to the second test set and the minimum word error rate is less than or equal to a tolerance threshold, determining, by the processing circuitry, that the second test set and the first test set satisfy the similarity condition.

### Claim 10 (depends on 1)

The method according to claim 1, further comprising:
obtaining voice data, segmenting the voice data into segments, and extracting a feature of each of the voice segments;
listing words corresponding to each of the voice segments;
converting the words into phonemes according to a pronunciation dictionary;
converting the phonemes into a hidden Markov model state according to a hidden Markov model;
obtaining a probability corresponding to each word according to a parameter of the hidden Markov model and a parameter of a Gaussian mixture model;
obtaining a hidden Markov model state sequence by comparing the probabilities; and
obtaining a duration of each hidden Markov model state according to the hidden Markov model state sequence.

### Claim 11

A computer device, comprising:
processing circuitry configured to
obtain a first test set of data selected from a first voice data set, and a first voice model parameter obtained by performing first voice model training based on the first voice data set;
randomly select data from a second voice data set to generate a second test set; and
perform second voice model training based on the second voice data set and the first voice model parameter when the second test set and the first test set satisfy a similarity condition.

### Claim 12 (depends on 11)

The computer device according to claim 11, wherein
the processing circuitry is configured to:
separately select data from the first voice data set to generate a training set and the first test set;
perform the first voice model training based on the training set to obtain a preset quantity of first voice models; and
test the first test set with the preset quantity of first voice models separately to obtain at least one first voice model having a word error rate within a preset range; wherein
the first voice model parameter is a parameter of one of the at least one first voice model having the word error rate within the preset range as the first voice model parameter.

### Claim 13 (depends on 12)

The computer device according to claim 12, wherein the processing circuitry is configured to:
randomly select a first preset proportion of data or a first fixed quantity of pieces of the data from the training set each time to perform the first voice model training, and repeating the first voice model training to obtain the preset quantity of first voice models.

### Claim 14 (depends on 12)

The computer device according to claim 12, wherein the processing circuitry is configured to:
test the first test set with the preset quantity of first voice models separately to obtain a word error rate of each of the first voice models; and
determine the at least one first voice model having the word error rate within the preset range according to the word error rate of each of the first voice models.

### Claim 15 (depends on 14)

The computer device according to claim 14, wherein the processing circuitry is configured to:
perform the second voice model training based on the second voice data set and the first voice model parameter of the first voice model having the word error rate being a minimum word error rate within the preset range.

### Claim 16 (depends on 15)

The computer device according to claim 15, wherein the processing circuitry is configured to:
test the second test set with the first voice model corresponding to the word error rate being the minimum word error rate within the preset range to obtain a word error rate corresponding to the second test set; and
when a difference between the word error rate corresponding to the second test set and the minimum word error rate within the preset range is less than or equal to a tolerance threshold, determine that the second test set and the first test set satisfy the similarity condition.

### Claim 17 (depends on 12)

The computer device according to claim 12, wherein the processing circuitry is configured to:
obtaining a ratio of a quantity of pieces of the data in the first test set to a quantity of pieces of data in the first voice data set, and randomly select data from the second voice data set according to the ratio to generate the second test set.

### Claim 18 (depends on 12)

The computer device according to claim 12, wherein the processing circuitry is configured to:
separately select data from the first voice data set to generate a training set and the first test set;
perform the first voice model training based on the training set to obtain a preset quantity of first voice models;
test the first test set with the preset quantity of first voice models separately to obtain a first voice model of the preset quantity of first voice models having a minimum word error rate; and
update the first voice model parameter based on a parameter of the first voice model having the minimum word error rate.

### Claim 19 (depends on 18)

The computer device according to claim 18, wherein the processing circuitry is configured to:
testing the second test set with the first voice model having the minimum word error rate within the preset quantity to obtain a word error rate corresponding to the second test set; and
when a difference between the word error rate corresponding to the second test set and the minimum word error rate is less than or equal to a tolerance threshold, determine that the second test set and the first test set satisfy the similarity condition.

### Claim 20

A non-transitory computer-readable storage medium, storing computer-readable instructions, the computer-readable instructions, when executed by one or more processors, causing the one or more processors to perform:
obtaining a first test set of data from a first voice data set, and a first voice model parameter obtained by performing first voice model training based on a first voice data set;
randomly selecting data from a second voice data set to generate a second test set; and
performing second voice model training based on the second voice data set and the first voice model parameter when the second test set and the first test set satisfy a similarity condition.
