# US 2010/0169090 A1 · Weighted sequential variance adaptation with prior knowledge for noise robust speech recognition

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

- Number: US 2010/0169090 A1
- Document ID: US20100169090A1
- Status: Published application
- Kind: A1
- Application number: 12/347,504
- Filed: 2008-12-31
- Published: 2010-07-01
- Earliest claimed: 2008-12-31
- Assignees: Texas Instruments Incorporated
- Inventors: Xiaodong Cui; Kaisheng Yao
- Web page: https://patentgrep.com/patent/US20100169090A1

## CPC

- 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: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING

## Abstract

A method for adapting acoustic models used for automatic speech recognition is provided. The method includes estimating noise in a portion of a speech signal, determining a first estimated variance scaling vector using an estimated 2-order polynomial and the noise estimation, wherein the estimated 2-order polynomial represents a priori knowledge of a dependency of a variance scaling vector on noise, determining a second estimated variance scaling vector using statistics from prior portions of the speech signal, determining a variance scaling factor using the first estimated variance scaling vector and the second estimated variance scaling vector, and using the variance scaling factor to adapt an acoustic model.

## Claims

### Claim 1

A method for adapting acoustic models in a speech recognition system, the method comprising:
estimating noise in a portion of a speech signal;
determining a first estimated variance scaling vector using an estimated 2-order polynomial and the noise estimation, wherein the estimated 2-order polynomial represents a priori knowledge of a dependency of a variance scaling vector on noise;
determining a second estimated variance scaling vector using statistics from prior portions of the speech signal;
determining a variance scaling factor using the first estimated variance scaling vector and the second estimated variance scaling vector; and
using the variance scaling factor to adapt an acoustic model.

### Claim 2 (depends on 1)

The method of claim 1, further comprising selecting an empirically determined weighting factor, wherein the determining a variance scaling factor further comprises using the weighting factor in the determination of the variance scaling factor.

### Claim 3 (depends on 2)

The method of claim 2, wherein determining a variance scaling factor further comprises computing the sum of the first estimated variance scaling vector multiplied by the weighting factor and the second estimated variance scaling factor multiplied by the weighting factor subtracted from the number one.

### Claim 4 (depends on 2)

The method of claim 2, wherein the weighting factor is selected based on an amount of change in noise level across previous portions of the speech signal.

### Claim 5 (depends on 4)

The method of claim 4, wherein a first weighting factor is selected when the amount of change is less than a predetermined threshold and a second weighting factor is selected when the amount of change exceeds the predetermined threshold.

### Claim 6 (depends on 5)

The method of claim 5, wherein the first weighting factor is 0.2 and the second weighting factor is 0.6.

### Claim 7 (depends on 1)

The method of claim 1, wherein the estimated 2-order polynomial is obtained by minimum square estimation of a 2-order polynomial regression on a plurality of variance scaling vectors, wherein the plurality of variance scaling vectors are estimated from speech signals collected under noisy conditions, wherein estimation of a variance scaling vector for a portion of a speech signal is performed using statistics from prior portions of the speech signal.

### Claim 8 (depends on 1)

The method of claim 1, wherein the portion of the speech signal comprises an utterance.

### Claim 9 (depends on 1)

The method of claim 1 wherein determining the second estimated variance scaling vector further comprises computing
ρ
-
∑
mt

γ
m

(
t
)

[
-
1
2
+
1
2

(
Y

(
t
)
-
μ
)
2

Σ
-
1

exp

(
-
ρ
)
]
∑
mt

γ
m

(
t
)

[
-
1
2

(
Y

(
t
)
-
μ
)
2

Σ
-
1

exp

(
-
ρ
)
]
where ρ is a previous estimate of the second estimated scaling vector, γm(t) is the posterior probability of state m at time t, μ denotes the mean of an acoustic model, and Σ denotes a variance vector of the acoustic model.

### Claim 10 (depends on 1)

The method of claim 1, wherein determining the first estimated variance scaling vector further comprises computing
âN2+{circumflex over (b)}N+ĉ
where N is the noise estimation, and â, {circumflex over (b)}, and ĉ are regression coefficients from a mean square estimation of a 2-order polynomial regression determined from speech signals collected under noisy conditions.

### Claim 11

A speech recognition system embodied in a digital system, the speech recognition system comprising:
a microphone;
a plurality of acoustic models stored in a memory of the digital system;
a speech recognizer operatively connected to the microphone to receive a speech signal and configured to use the acoustic models to recognize speech in the speech signal; and
a model adaptor subsystem operatively connected to the microphone to receive the speech signal and configured to adapt at least one acoustic model of the plurality of acoustic models by
estimating noise in a portion of the speech signal;
determining a first estimated variance scaling vector using an estimated 2-order polynomial and the noise estimation, wherein the estimated 2-order polynomial represents a priori knowledge of a dependency of a variance scaling vector on noise;
determining a second estimated variance scaling vector using statistics from prior portions of the speech signal;
determining a variance scaling factor using the first estimated variance scaling vector and the second estimated variance scaling vector; and
using the variance scaling factor to adapt the at least one acoustic model.

### Claim 12 (depends on 11)

The speech recognition system of claim 11, wherein the model adaptor subsystem is further configured to select an empirically determined weighting factor, wherein the determining a variance scaling factor further comprises using the weighting factor in the determination of the variance scaling factor.

### Claim 13 (depends on 12)

The speech recognition system of claim 12, wherein the weighting factor is selected based on an amount of change in noise level across previous portions of the speech signal.

### Claim 14 (depends on 13)

The speech recognition system of claim 13, wherein a first weighting factor is selected when the amount of change is less than a predetermined threshold and a second weighting factor is selected when the amount of change exceeds the predetermined threshold.

### Claim 15 (depends on 14)

The speech recognition system of claim 14, wherein the first weighting factor is 0.2 and the second weighting factor is 0.6.

### Claim 16 (depends on 11)

The speech recognition system of claim 11, wherein the estimated 2-order polynomial is obtained by minimum square estimation of a 2-order polynomial regression on a plurality of variance scaling vectors, wherein the plurality of variance scaling vectors are estimated from speech signals collected under noisy conditions, wherein estimation of a variance scaling vector for a portion of a speech signal is performed using statistics from prior portions of the speech signal.

### Claim 17 (depends on 11)

The speech recognition system of claim 11, wherein the portion of the speech signal comprises an utterance.

### Claim 18 (depends on 11)

The speech recognition system of claim 11, wherein the digital system is comprised in a mobile device.

### Claim 19

A digital system comprising:
a microphone;
a memory configured to store a plurality of acoustic models and a plurality of instructions; and
a processor operatively coupled to the memory and configured to receive a portion of a speech signal from the microphone and to execute the plurality of instructions to perform a method for adapting at least one acoustic model of the plurality of acoustic models, the method comprising:
estimating noise in the portion of the speech signal;
determining a first estimated variance scaling vector using an estimated 2-order polynomial and the noise estimation, wherein the estimated 2-order polynomial represents a priori knowledge of a dependency of a variance scaling vector on noise;
determining a second estimated variance scaling vector using statistics from prior portions of the speech signal;
determining a variance scaling factor using the first estimated variance scaling vector and the second estimated variance scaling vector; and
using the variance scaling factor to adapt the at least one acoustic model.

### Claim 20 (depends on 19)

The digital system of claim 19, wherein the method further comprises selecting an empirically determined weighting factor, wherein the determining a variance scaling factor further comprises using the weighting factor in the determination of the variance scaling factor.