Method and system of automatic context-bound domain-specific speech recognition

Patent metadata

Field	Value
Publication number	US-12555572-B2
Application number	US-202117561849-A
Country	US
Kind code	B2
Filing date	Dec 24, 2021
Priority date	Dec 24, 2021
Publication date	Feb 17, 2026
Grant date	Feb 17, 2026

Abstract

A system, article, and method of automatic context-bound domain-specific speech recognition uses general language models.

First claim

What is claimed is:

1 . A computer-implemented method of audio processing, comprising:
obtaining an automatic speech recognition general dataset of general sentences or general phrases or both; generating a domain dataset comprising selecting at least portions of the general sentences or the general phrases or both with one or more domain phrases from a domain list, and adding the selected at least portions of the general sentences or the general phrases or both to the domain dataset; generating, by a processor, context n-grams to add to the domain dataset, including:
selecting n-grams in the selected general sentences or general phrases with the domain phrases,
selecting, for the domain dataset, multiple context n-grams wherein a position of the domain phrases varies within the general sentences or the general phrases or both,
wherein generating the context n-grams includes using a sliding window having a selected length, and wherein the selected length of the sliding window can be one or more of a fixed length and a varying length, and
based on a respective length of each of the domain phrases, determining whether the selected length of the sliding window is the fixed length or the varying length; and
training a domain language model arranged to recognize the domain phrases, and at least portions of domain sentences comprising using the domain dataset.

2 . The method of claim 1 wherein the domain is toxic language deemed to be undesired or inappropriate language in an environment present when audio is captured to be analyzed by the trained domain language model.

3 . The method of claim 1 wherein the at least portions of the general sentences or general phrases to be added to the domain dataset are entire sentences or entire phrases from the general dataset.

4 . The method of claim 1 wherein the generating context n-grams comprises selecting multiple context n-grams, each of the multiple context n-grams including a same domain word from a single general sentence or single general phrase, wherein the position of the same domain word varies within each context n-gram of the same domain word in the same general sentence or general phrase.

5 . The method of claim 4 wherein the sliding window has a fixed length and the context n-gram size in total number of words remains fixed.

6 . The method of claim 1 wherein the context n-grams are each three or five words.

7 . The method of claim 1 wherein generating context n-grams comprises selecting multiple context n-grams each of multiple words including a same domain word from a same position in one of the general sentences or general phrases, wherein the number of words from the general sentence or general phrase in each context n-gram is different of the multiple context n-grams.

8 . The method of claim 1 comprising adding the context n-grams of the general sentences or general phrases to the domain dataset instead of adding the selected general sentence and general phrase to the domain dataset associated with the context n-grams.

9 . The method of claim 1 comprising training the general language model; and forming a trained general and domain-specific language model comprising combining the general language model and the domain language model.

10 . The method of claim 1 , wherein generating context n-grams to add to the domain dataset further comprises determining that the selected length of the sliding window is the varying length based on the respective length of a respective domain phrase of the domain phrases.

11 . The method of claim 1 , wherein generating context n-grams to add to the domain dataset further comprises determining that the selected length of the sliding window is the varying length based on determining that a plurality of domain words or of the domain phrases are present in series in a selected general sentence or general phrase.

12 . A computer-implemented system of automatic domain speech recognition comprising:
memory storing an audio signal of human speech; and processor circuitry forming at least one processor communicatively connected to the memory, the at least one processor being arranged to operate by:
obtaining an automatic speech recognition general dataset of general sentences or general phrases or both;
generating a domain dataset comprising selecting at least portions of the general sentences or the general phrases or both with one or more domain phrases from a domain list, and adding the selected at least portions of the general sentences or the general phrases or both to the domain dataset;
generating context n-grams to add to the domain dataset, including:
selecting n-grams in the selected general sentences or general phrases with the domain phrases,
selecting, for the domain dataset, multiple context n-grams wherein a position of the domain phrases varies within the general sentences or the general phrases or both,
wherein generating the context n-grams includes using a sliding window having a selected length, and wherein the selected length of the sliding window can be one or more of a fixed length and a varying length, and
determining, based on a respective length of each of the domain phrases, whether the selected length of the sliding window is the fixed length or the varying length; and

training a domain language model arranged to recognize the domain phrases, and at least portions of domain sentences comprising using the domain dataset.

13 . The system of claim 12 wherein the at least portions of the general sentences or general phrases to be added to the domain dataset are entire sentences or entire phrases from the general dataset.

14 . The system of claim 12 wherein the generating context n-grams comprises selecting multiple context n-grams each of the multiple context n-grams including a same domain word from a single general sentence or single general phrase, wherein the position of the same domain word varies within each context n-gram of the same domain word in the same general sentence or general phrase.

15 . The system of claim 14 wherein the sliding window has a fixed length and the context n-gram size in total number of words remains fixed.

16 . The system of claim 12 wherein the at least one processor is arranged to operate by training the general language model; and forming a trained general and domain-specific language model comprising combining the general language model and the domain language model.

17 . The system of claim 12 , wherein generating context n-grams to add to the domain dataset further comprises determining that the selected length of the sliding window is the varying length based on the respective length of a respective domain phrase of the domain phrases.

18 . The system of claim 12 , wherein generating context n-grams to add to the domain dataset further comprises determining that the selected length of the sliding window is the varying length based on determining that a plurality of domain words or of the domain phrases are present in series in a selected general sentence or general phrase.

19 . At least one non-transitory computer readable medium comprising a plurality of instructions that in response to being executed on a computing device, causes the computing device to operate by:
obtaining an automatic speech recognition general dataset of general sentences or general phrases or both; generating a domain dataset comprising selecting at least portions of the general sentences or the general phrases or both with one or more domain phrases from a doma

Assignees

• Intel Corp

Inventors

• Jessa Szymon
• Nowicki Jakub
• Papaj Michal
• Hoffmann Piotr
• Swider Krzysztof
• Stemmer Georg

Classifications

• G10L15/063 Primary

  Training · CPC title

• G10L2015/025

  Phonemes, fenemes or fenones being the recognition units · CPC title

• G10L2015/0631

  Creating reference templates; Clustering · CPC title

• G10L25/30

  using neural networks · CPC title

• G10L25/51

  for comparison or discrimination · CPC title