# US 2020/0265833 A1 · Best path change rate for unsupervised language model weight selection

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

- Number: US 2020/0265833 A1
- Document ID: US20200265833A1
- Status: Published application
- Kind: A1
- Application number: 16/279,491
- Filed: 2019-02-19
- Published: 2020-08-20
- Earliest claimed: 2019-02-19
- Assignees: Tencent America LLC
- Applicants: Tencent America LLC
- Inventors: Peidong WANG; Jai CUI; Chao WENG; Dong YU
- Web page: https://patentgrep.com/patent/US20200265833A1

## CPC

- G10L 15/183: using context dependencies, e.g. language models
- G10L 15/063: {Training}
- G10L 25/03: characterised by the type of extracted parameters
- G10L 25/27: characterised by the analysis technique
- 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 selecting an optimal language model weight (LMW) used to perform automatic speech recognition, including decoding test audio into a lattice using a language model; analyzing the lattice using a first LMW of a plurality of LMWs to determine a first plurality of best paths; analyzing the lattice using a second LMW of the plurality of LMWs to determine a second plurality of best paths; determining a first best path change rate (BCPR) corresponding to the first LMW based on a number of best path changes between the first plurality of best paths and the second plurality of best paths; and determining the first LMW to be the optimal LMW based on the first BCPR being a lowest BCPR from among a plurality of BCPRs corresponding to the plurality of LMWs.

## Claims

### Claim 1

A method for selecting an optimal language model weight (LMW) used to perform automatic speech recognition, the method comprising:
decoding test audio into a lattice using a language model;
analyzing the lattice using a first LMW of a plurality of LMWs to determine a first plurality of best paths;
analyzing the lattice using a second LMW of the plurality of LMWs to determine a second plurality of best paths;
determining a first best path change rate (BCPR) corresponding to the first LMW based on a number of best path changes between the first plurality of best paths and the second plurality of best paths; and
determining the first LMW to be the optimal LMW based on the first BCPR being a lowest BCPR from among a plurality of BCPRs corresponding to the plurality of LMWs.

### Claim 2 (depends on 1)

The method of claim 1, wherein the plurality of LMWs comprise a plurality of scaling factors used to combine the language model with an acoustic model.

### Claim 3 (depends on 1)

The method of claim 1, wherein each LMW of the plurality of LMWs is separated from neighboring LMWs of the plurality of LMWs by a fixed step size.

### Claim 4 (depends on 1)

The method of claim 1, wherein each score of a plurality of scores corresponding to the first plurality of best paths comprises a combination of an acoustic score and a language model score balanced by the first LMW, and wherein each best path of the first plurality of best paths comprises a path in the lattice with a largest score.

### Claim 5 (depends on 1)

The method of claim 1, wherein the number of the best path changes is represented as follows:
c

(
w
)
=
∑
s
∈
S

δ

(
p
s

(
w
)
≠
p
s

(
w
+
ϵ
)
)
,
wherein c(w) denotes the number of the best path changes, s is an index of an utterance, S denotes utterances in a data set corresponding to the test audio, δ( ) denotes a Kronecker delta function, ps( )denotes a best path, w denotes the first LMW, ε denotes a fixed step size, and w+ε denotes the second LMW.

### Claim 6 (depends on 5)

The method of claim 5, wherein the first BCPR is represented as follows:
r

(
w
)
=
c

(
w
)
N
,
wherein r(w) denotes the first BCPR and N denotes a total number of the utterances in the data set.

### Claim 7 (depends on 6)

The method of claim 6, wherein a continuous function corresponding to the BCPR is represented as follows:
R(wc)=Ir(wc)+Dr(wc),
wherein R(wc) denotes the continuous function corresponding to the BCPR, wherein Ir(wc) denotes a normalized number of best paths which increase a word error rate after a LMW change, and Dr(wc) denotes a normalized number of best paths which decrease the word error rate after the LMW change.

### Claim 8 (depends on 7)

The method of claim 7, wherein Ir(wc) and Dr(wc) are represented as follows:
Ir(wc)|wc∈[wopt−ε]=−k*wc+bI,
and
Dr(wc)|wc∈[wopt−ε,wopt+ε]=k*wc+bD,
wherein wopt denotes the optimal LMW, k denotes (k≥0) a slope, bI denotes an intercept for Ir(wc), and bD denotes an intercept for Dr(wc).

### Claim 9 (depends on 8)

The method of claim 8, wherein R(wc) is represented as follows:
R(wc)|wc∈[wopt−ε,woptε]=Ir(wc)+Dr(wc)=bI+bD,
and
wherein R′(wopt) is represented as follows:
R
′

(
w
opt
)
=
dR

(
w
c
)
dw
c

|
w
opt
=
0.

### Claim 10 (depends on 9)

The method of claim 9, wherein the first LMW is determined to be the optimal LMW based on the first BCPR being closest to a minimum value of R′(wopt) from among the plurality of BCPRs.

### Claim 11

A device for selecting an optimal language model weight (LMW) used to perform automatic speech recognition, the device comprising:
at least one memory configured to store program code; and
at least one processor configured to read the program code and operate as instructed by the program code, the program code including:
decoding code configured to cause the at least one processor to decode test audio into a lattice using a language model,
first analyzing code configured to cause the at least one processor to analyze the lattice using a first LMW of a plurality of LMWs to determine a first plurality of best paths,
second analyzing code configured to cause the at least one processor to analyze the lattice using a second LMW of the plurality of LMWs to determine a second plurality of best paths,
first determining code configured to cause the at least one processor to determine a first best path change rate (BCPR) corresponding to the first LMW based on a number of best path changes between the first plurality of best paths and the second plurality of best paths, and
second determining code configured to cause the at least one processor to determine the first LMW to be the optimal LMW based on the first BCPR being a lowest BCPR from among a plurality of BCPRs corresponding to the plurality of LMWs.

### Claim 12 (depends on 11)

The device of claim 11, wherein the plurality of LMWs comprise a plurality of scaling factors used to combine the language model with an acoustic model.

### Claim 13 (depends on 11)

The device of claim 11, wherein each LMW of the plurality of LMWs is separated from neighboring LMWs of the plurality of LMWs by a fixed step size.

### Claim 14 (depends on 11)

The device of claim 11, wherein the number of the best path changes is represented as follows:
c

(
w
)
=
∑
s
∈
S

δ

(
p
s

(
w
)
≠
p
s

(
w
+
ϵ
)
)
,
wherein c(w) denotes the number of the best path changes, s is an index of an utterance, S denotes utterances in a data set corresponding to the test audio, δ( ) denotes a Kronecker delta function, ps( )denotes a best path, w denotes the first LMW, ε denotes a fixed step size, and w+ε denotes the second LMW.

### Claim 15 (depends on 14)

The device of claim 14, wherein the first BCPR is represented as follows:
r

(
w
)
=
c

(
w
)
N
,
wherein r(w) denotes the first BCPR and N denotes a total number of the utterances in the data set.

### Claim 16 (depends on 15)

The device of claim 15, wherein a continuous function corresponding to the BCPR is represented as follows:
wherein R(wc) denotes the continuous function corresponding to the BCPR, wherein Ir(wc) denotes a normalized number of best paths which increase a word error rate after a LMW change, and Dr(wc) denotes a normalized number of best paths which decrease the word error rate after the LMW change.

### Claim 17 (depends on 16)

The device of claim 16, wherein Ir(wc) and Dr(wc) are represented as follows:
Ir(wc)|wc∈[wopt−ε,wopt+ε]=−k*wc+bI,
and
Dr(wc)|wc∈[wopt−ε,wopt+ε]=k*wc+bD,
wherein wopt denotes the optimal LMW, k denotes (k≥0) a slope, bI denotes an intercept for Ir(wc), and bD denotes an intercept for Dr(wc).

### Claim 18 (depends on 17)

The device of claim 17, wherein R(wc) is represented as follows:
R(wc)|wc∈[wopt−ε,wopt+ε]=Ir(wc)+Dr(wc)=bI+bD,
and
wherein R′(wopt) is represented as follows:
R
′

(
w
opt
)
=
dR

(
w
c
)
dw
c

|
w
opt
=
0.

### Claim 19 (depends on 18)

The device of claim 18, wherein the first LMW is determined to be the optimal LMW based on the first BCPR being closest to a minimum value of R′(wopt) from among the plurality of BCPRs.

### Claim 20

A non-transitory computer-readable medium storing instructions, the instructions comprising: one or more instructions that, when executed by one or more processors of a device for selecting an optimal language model weight (LMW) used to perform automatic speech recognition, cause the one or more processors to:
decode test audio into a lattice using a language model,
analyze the lattice using a first LMW of a plurality of LMWs to determine a first plurality of best paths,
analyze the lattice using a second LMW of the plurality of LMWs to determine a second plurality of best paths,
determine a first best path change rate (BCPR) corresponding to the first LMW based on a number of best path changes between the first plurality of best paths and the second plurality of best paths, and
determine the first LMW to be the optimal LMW based on the first BCPR being a lowest BCPR from among a plurality of BCPRs corresponding to the plurality of LMWs.
