# US 8,571,857 B2 · System and method for generating models for use in automatic speech recognition

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

- Number: US 8,571,857 B2
- Document ID: US8571857B2
- Status: Granted
- Kind: B2
- Application number: 12/908,222
- Filed: 2010-10-20
- Published: 2012-04-26
- Granted: 2013-10-29
- Earliest claimed: 2010-10-20
- Assignees: AT&T Intellectual Property I, L.P.
- Applicants: John Mocenigo; Mazin Gilbert; Heather Wighton
- Inventors: John Mocenigo; Mazin Gilbert; Heather Wighton
- Web page: https://patentgrep.com/patent/US8571857B2

## CPC

- G10L 15/063: {Training}
- G10L 2015/0638: {Interactive procedures}
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING

## Abstract

Disclosed herein are systems, methods, and non-transitory computer-readable storage media for generating a model for use with automatic speech recognition. These principles can be implemented as part of a streamlined tool for automatic training and tuning of speech, or other, models with a fast turnaround and with limited human involvement. A system configured to practice the method receives, as part of a request to generate a model, input data and a seed model. The system receives a cost function indicating accuracy and at least one of speed and memory usage, The system processes the input data based on seed model and based on parameters that optimize the cost function to yield an updated model, and outputs the updated model.

## Claims

### Claim 1

A method comprising:
receiving, from a first entity, a request to generate a model, wherein the request comprises input data and a seed model;
receiving a cost function associated with generation of the model, the cost function indicating an accuracy and one of speed and memory usage, wherein the cost function is formulated as:
Accscore(ASR(Xi))=f(Xi)→word accuracy and speed
where the following definitions apply:
fmin(Xi)=−1*(word accuracy−β*speed),
speed=(CPU time)/(audio time),
β=a weighting factor to speed that provides a tradeoff between accuracy and speed,
Xi=a set of parameters that affect accuracy and speed, such as beam width, LM scale, MAP multiplier, maximum active arcs, duration scale;
processing, at a second entity, the input data based on the seed model and based on parameters that modify one of the accuracy and the one of the speed and the memory usage of the cost function, to yield an updated model; and
outputting the updated model.

### Claim 2 (depends on 1)

The method of claim 1, wherein the input data comprises one of transcribed speech data, audio data, hex data, web data, and data stored in a database.

### Claim 3 (depends on 1)

The method of claim 1, wherein the seed model comprises one of an acoustic model, a language model, a probabilistic parsing model, a pronunciation model, a text normalization model, and a classification model.

### Claim 4 (depends on 1)

The method of claim 1, wherein processing the input data further comprises one of building the updated model, tuning the updated model, and certifying the updated model.

### Claim 5 (depends on 1)

The method of claim 1, wherein the parameters result in the cost function yielding a known output.

### Claim 6 (depends on 1)

The method of claim 1, wherein the updated model is ready for implementation in an automatic speech recognizer.

### Claim 7 (depends on 1)

The method of claim 1, wherein processing the input data further comprises iteratively applying an algorithm to the input data to determine settings for the updated model.

### Claim 8 (depends on 1)

The method of claim 1, wherein the parameters comprise one of beam width, grammar scale, word insertion penalty, maximum arc length, maximum number of arcs allowed at any point in time, and state duration.

### Claim 9 (depends on 1)

The method of claim 1, wherein processing the input data comprises a first pass for retraining, a second pass for vocal-tract length normalization, and a third pass for constrained model adaptation.

### Claim 10

A system comprising:
a processor; and
a computer-readable storage medium having instructions stored which, when executed by the processor, result in the processor performing operations comprising:
receiving, from a separate entity, a request to generate a model, input data and a seed model;
receiving a cost function associated with generation of the model, the cost function indicating an accuracy and one of speed and memory usage, wherein the cost function is formulated as:
Accscore(ASR(Xi))=f(Xi)→word accuracy and speed
where the following definitions apply:
fmin(Xi)=−1* (word accuracy−β*speed),
speed=(CPU time)/(audio time),
β=a weighting factor to speed that provides a tradeoff between accuracy and speed
Xi=a set of parameters that affect accuracy and speed, such as beam width, LM scale, MAP multiplier, maximum active arcs, duration scale;
processing the input data based on the seed model and based on parameters that modify one of the accuracy and the one of speed and memory of the cost function, to yield an updated model; and
outputting the updated model.

### Claim 11 (depends on 10)

The system of claim 10, wherein the input data comprises one of transcribed speech data, untranscribed speech data, audio data, hex data, web data, and data stored in a database.

### Claim 12 (depends on 10)

The system of claim 10, wherein the seed model comprises one of an acoustic model, a language model, a probabilistic parsing model, a pronunciation model, a text normalization model, and a classification model.

### Claim 13 (depends on 10)

The system of claim 10, wherein processing the input data further comprises building the updated model, tuning the updated model, and certifying the updated model.

### Claim 14 (depends on 10)

The system of claim 10, wherein the parameters result in the cost function yielding a known output.

### Claim 15 (depends on 10)

The system of claim 10, wherein the updated model is ready for implementation in an automatic speech recognizer.

### Claim 16

A computer-readable storage device having instructions stored which, when executed by a computing device, cause the computing device to perform operations comprising:
receiving, from a separate entity, of a request to generate a model, input data and a seed model;
receiving a cost function associated with generation of the model, the cost function indicating accuracy and one of speed and memory usage, wherein the cost function is formulated as:
Accscore(ASR(Xi))=f(Xi)→word accuracy and speed
where the following definitions apply:
fmin(Xi)=−1* (word accuracy−β*speed),
speed=(CPU time)/(audio time),
β=a weighting factor to speed that provides a tradeoff between accuracy and speed,
Xi=a set of parameters that affect accuracy and speed, such as beam width, LM scale, MAP multiplier, maximum active arcs, duration scale;
processing the input data based on the seed model and based on parameters that modify the accuracy and the one of speed and memory usage of the cost function, to yield an updated model; and
outputting the updated model.

### Claim 17 (depends on 16)

The computer-readable storage device of claim 16, wherein the updated model is ready for implementation in an automatic speech recognizer.

### Claim 18 (depends on 16)

The computer-readable storage device of claim 16, wherein processing the input data further comprises iteratively applying an algorithm to the input data to determine settings for the updated model.

### Claim 19 (depends on 16)

The computer-readable storage device of claim 16, wherein the parameters comprise one of beam width, grammar scale, word insertion penalty, maximum arc length, maximum number of arcs allowed at any point in time, and state duration.

### Claim 20 (depends on 16)

The computer-readable storage device of claim 16, wherein processing the input data comprises a first pass for retraining, a second pass for vocal-tract length normalization, and a third pass for constrained model adaptation.