# US 10,373,609 B2 · Voice recognition method and apparatus

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

- Number: US 10,373,609 B2
- Document ID: US10373609B2
- Status: Granted
- Kind: B2
- Application number: 15/619,252
- Filed: 2017-06-09
- Published: 2018-07-05
- Granted: 2019-08-06
- Earliest claimed: 2016-12-29
- Assignees: Baidu Online Network Technology (Beijing) Co., Ltd.
- Applicants: Baidu Online Network Technology (Beijing) Co., Ltd.
- Inventors: Weiwei Cui; Yu Sun
- Web page: https://patentgrep.com/patent/US10373609B2

## CPC

- G10L 15/063: {Training}
- G10L 15/02: Feature extraction for speech recognition; Selection of recognition unit
- G10L 15/16: using artificial neural networks
- G10L 15/20: Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech (G10L21/02 takes precedence)
- G10L 2021/02082: {the noise being echo, reverberation of the speech}
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING

## Abstract

The present application discloses a voice recognition method and apparatus. A specific implementation of the method includes: in response to detecting a microphone receiving voice signal containing interfering sound signal, performing high-pass filtering on the voice signal; cancelling the interfering sound signal in the voice signal subjected to high-pass filtering; performing automatic gain control on the voice signal subjected to cancelling the interfering sound signal, to obtain target voice signal; and extracting a feature vector from the target voice signal and inputting the feature vector into a pre-trained acoustic model, to obtain a voice recognition result matching the target voice signal, the acoustic model being used for representing a corresponding relationship between the feature vector and the voice recognition result. This implementation improves the success rate of voice recognition.

## Claims

### Claim 1

A voice recognition method for a terminal device, the terminal device being equipped with a microphone, the method comprising:
performing high-pass filtering on a voice signal, in response to detecting the microphone receiving the voice signal containing interfering sound signal;
cancelling the interfering sound signal in the voice signal subjected to high-pass filtering;
performing automatic gain control on the voice signal subjected to cancelling the interfering sound signal, to obtain target voice signal; and
extracting a feature vector from the target voice signal and inputting the feature vector into a pre-trained acoustic model, to obtain a voice recognition result matching the target voice signal, the acoustic model being used for representing a corresponding relationship between the feature vector and the voice recognition result, wherein before the performing high-pass filtering on the voice signals, in response to detecting the microphone receiving the voice signal, the method further comprises:
pre-processing a pre-acquired training sample to generate a target training sample, the target training sample comprising a voice identifier;
extracting the feature vector from the target training sample; and
training, using a convolutional neural network, a deep neural network, and a restricted Boltzmann machine, and assigning the feature vector extracted from the target training sample as an input and the voice identifier as an output, to obtain the acoustic model.

### Claim 2 (depends on 1)

The voice recognition method according to claim 1, wherein the terminal device is further equipped with a loudspeaker, the interfering sound signal comprises an echo signal and a noise signal, and the echo signal is a sound signal sent by the loudspeaker and transmitted to the microphone.

### Claim 3 (depends on 2)

The voice recognition method according to claim 2, wherein the cancelling the interfering sound signal in the voice signal to obtain target voice signal comprises:
performing adaptive filtering on the voice signal subjected to high-pass filtering using a time delay estimation algorithm, to cancel the echo signal; and
cancelling the noise signal in the voice signal subjected to adaptive filtering using a noise suppression algorithm.

### Claim 4 (depends on 1)

The voice recognition method according to claim 1, wherein the pre-processing a pre-acquired training sample to generate a target training sample comprises:
performing high-pass filtering on the pre-acquired training sample;
performing sequentially echo cancellation and noise suppression on the training sample subjected to high-pass filtering; and
performing automatic gain control on the training sample subjected to noise suppression, to generate the target training sample.

### Claim 5 (depends on 1)

The voice recognition method according to claim 1, wherein before the performing high-pass filtering on the voice signals, in response to detecting the microphone receiving a voice signal, the method further comprises:
clustering the voice identifier outputted by the acoustic model using a clustering algorithm, and determining voice identifier obtained after clustering as a voice recognition result matching the training sample.

### Claim 6

A voice recognition apparatus for a terminal device, the terminal device being equipped with a microphone, the apparatus comprising:
at least one processor; and
a memory storing instructions, which when executed by the at least one processor, cause the at least one processor to perform operations, the operations comprising:
performing high-pass filtering on the voice signal, in response to detecting the microphone receiving voice signal containing interfering sound signal;
cancelling the interfering sound signal in the voice signal subjected to high-pass filtering;
performing automatic gain control on the voice signal subjected to cancelling the interfering sound signal, to obtain target voice signal; and
extracting a feature vector from the target voice signal and input the feature vector into a pre-trained acoustic model, to obtain a voice recognition result matching the target voice signal, the acoustic model being used for representing a corresponding relationship between the feature vector and the voice recognition result, wherein before the performing high-pass filtering on the voice signals, in response to detecting the microphone receiving the voice signal, the operations further comprise:
pre-processing a pre-acquired training sample to generate a target training sample, the target training sample comprising a voice identifier;
extracting the feature vector from the target training sample; and
training, using a convolutional neural network, a deep neural network, and a restricted Boltzmann machine, and assigning the feature vector extracted from the target training sample as an input and the voice identifier as an output, to obtain the acoustic model.

### Claim 7 (depends on 6)

The voice recognition apparatus according to claim 6, wherein the terminal device is further equipped with a loudspeaker, the interfering sound signal comprises an echo signal and a noise signal, and the echo signal is a sound signal sent by the loudspeaker and transmitted to the microphone.

### Claim 8 (depends on 7)

The voice recognition apparatus according to claim 7, wherein the cancelling the interfering sound signal in the voice signal to obtain target voice signal comprises:
performing adaptive filtering on the voice signal subjected to high-pass filtering using a time delay estimation algorithm, to cancel the echo signal; and
cancelling the noise signal in the voice signal subjected to adaptive filtering using a noise suppression algorithm.

### Claim 9 (depends on 6)

The voice recognition apparatus according to claim 6, wherein the pre-processing a pre-acquired training sample to generate a target training sample comprises:
performing high-pass filtering on the pre-acquired training sample;
performing sequentially echo cancellation and noise suppression on the training sample subjected to high-pass filtering; and
performing automatic gain control on the training sample subjected to noise suppression, to generate the target training sample.

### Claim 10 (depends on 6)

The voice recognition apparatus according to claim 6, wherein the before the performing high-pass filtering on the voice signals, in response to detecting the microphone receiving a voice signal, the operations further comprise:
clustering the voice identifier outputted by the acoustic model using a clustering algorithm, and determine a voice identifier obtained after clustering as a voice recognition result matching the training sample.

### Claim 11

A non-transitory computer storage medium storing a computer program, which when executed by one or more processors, cause the one or more processors to perform operations, the operations comprising:
performing high-pass filtering on a voice signal, in response to detecting the microphone receiving the voice signal containing interfering sound signal;
cancelling the interfering sound signal in the voice signal subjected to high-pass filtering;
performing automatic gain control on the voice signal subjected to cancelling the interfering sound signal, to obtain target voice signal; and
extracting a feature vector from the target voice signal and inputting the feature vector into a pre-trained acoustic model, to obtain a voice recognition result matching the target voice signal, the acoustic model being used for representing a corresponding relationship between the feature vector and the voice recognition result, wherein before the performing high-pass filtering on the voice signals, in response to detecting the microphone receiving the voice signal, the operations further comprise:
pre-processing a pre-acquired training sample to generate a target training sample, the target training sample comprising a voice identifier;
extracting the feature vector from the target training sample; and
training, using a convolutional neural network, a deep neural network, and a restricted Boltzmann machine, and assigning the feature vector extracted from the target training sample as an input and the voice identifier as an output, to obtain the acoustic model.

### Claim 12 (depends on 11)

The non-transitory computer storage medium according to claim 11, wherein the terminal device is further equipped with a loudspeaker, the interfering sound signal comprises an echo signal and a noise signal, and the echo signal is a sound signal sent by the loudspeaker and transmitted to the microphone.

### Claim 13 (depends on 12)

The non-transitory computer storage medium according to claim 12, wherein the cancelling the interfering sound signal in the voice signal to obtain target voice signal comprises:
performing adaptive filtering on the voice signal subjected to high-pass filtering using a time delay estimation algorithm, to cancel the echo signal; and
cancelling the noise signal in the voice signal subjected to adaptive filtering using a noise suppression algorithm.

### Claim 14 (depends on 11)

The non-transitory computer storage medium according to claim 11, wherein the pre-processing a pre-acquired training sample to generate a target training sample comprises:
performing high-pass filtering on the pre-acquired training sample;
performing sequentially echo cancellation and noise suppression on the training sample subjected to high-pass filtering; and
performing automatic gain control on the training sample subjected to noise suppression, to generate the target training sample.

### Claim 15 (depends on 11)

The non-transitory computer storage medium according to claim 11, wherein before the performing high-pass filtering on the voice signals, in response to detecting the microphone receiving a voice signal, the operations further comprise:
clustering the voice identifier outputted by the acoustic model using a clustering algorithm, and determining voice identifier obtained after clustering as a voice recognition result matching the training sample.