- Number
- 12633290
- Published
- 2026-05-19
- Filed
- 2022-02-07
- Assignee
- Disney Enterprises, Inc.
- Inventors
- Tiwari; Sanchita, Yu; Xiuyang, Kennedy; Justin Ali, Van Dall; Dirk
- CPC
- G06F40/279; G10L15/26; G06N3/0442; G06N3/0455; G06N3/088; G06N3/09; G06N7/01; G06N20/00; G10L15/02; G10L15/16; G10L15/1815; G10L2015/025
- Verdict
- Set aside streaming, nlp/localization
- Source
- Google Patents · FreePatentsOnline
The keeper's note
A system includes a computing platform including processing hardware and a system memory storing software code.
Abstract
A system includes a computing platform including processing hardware and a system memory storing software code. The processing hardware executes the software code to receive a transcript of speech by a user, generate a phoneme stream corresponding to the transcript, partition the phoneme stream into words, and aggregate subsets of the words to form candidate sentences. The software code is further executed to determine, using one or both of an entity identified in the transcript and a history of the user, one or both of a user intent and a context of the speech by the user, rank the candidate sentences and the transcript based on one or both of the user intent and the context of the speech by the user, and identify, based on the ranking, one of the candidate sentences or the transcript as the best transcription of the speech by the user.
Background
BACKGROUND (1) Most modern automatic speech recognition (ASR) systems are capable of high quality performance under ideal conditions, such as low background noise and clearly enunciated, unaccented human speech. However, interpretive errors can occur when ASR systems are used in noisy environments, or when speech is indistinctly uttered due to age or disability, is heavily accented, is of a different vernacular, or includes linguistic mistakes. Moreover, the reliability of ASR systems can also rely on the corpus of words used to train the system. Consequently, an ASR system trained on a corpus of common language terms may be unreliable when used to transcribe speech that is unique to or characteristic of a specialized language domain, such as medical or other scientific terminology, for example. Consequently, there remains a need in the art for a natural language understanding (NLU) solution capable of performing domain and user intent specific disambiguation of transcribed speech.
Claims
1. A system comprising: a computing platform including a processing hardware and a system memory storing a software code, a trained common language prediction machine learning (ML) model, a trained domain-specific language prediction ML model, and a trained word association ML model; the processing hardware configured to execute the software code to: receive a transcript of speech by a user; generate a phoneme stream corresponding to the transcript; partition the phoneme stream into a plurality of words; randomly generate, using the plurality of words partitioned from the phoneme stream, a plurality of word sequences of variable length; identify, using dynamic time warping (DTW) applied to the randomly generated word sequences of variable length, a candidate word pool; designate, using the trained common language prediction ML model, the trained domain-specific language prediction ML model and the trained word association ML model to predict a disambiguation of the transcript and each of a plurality of different sequences of words included in the candidate word pool, a plurality of candidate sentences, wherein designating the plurality of candidate sentences further includes determining a confidence score for the transcript and each of the plurality of candidate sentences; determine, using at least one of an entity identified in the transcript or a history of the user, at least one of an intent of the user or a context of the speech by the user; perform a ranking of the plurality of candidate sentences and the transcript based on the at least one of the intent of the user or the context of the speech by the user; and identify, based on the ranking, one of the plurality of candidate sentences or the transcript as a best transcription of the speech by the user. ||
10. A method for use by a system including a computing platform having processing hardware and a system memory storing a software code, a trained common language prediction machine learning (ML) model, a trained domain-specific language prediction ML model, and a trained word association ML model, the method comprising: receiving, by the software code executed by the processing hardware, a transcript of speech by a user; generating, by the software code executed by the processing hardware, a phoneme stream corresponding to the transcript; partitioning, by the software code executed by the processing hardware, the phoneme stream into a plurality of words; randomly generating, by the software code executed by the processing hardware and using the plurality of words partitioned from the phoneme stream, a plurality of word sequences of variable length; identifying, by the software code executed by the processing hardware and using dynamic time warping (DTW) applied to the randomly generated word sequences of variable length, a candidate word pool; designating, by the software code executed by the processing hardware and using the trained common language prediction ML model, the trained domain-specific language prediction ML model and the trained word association ML model to predict a disambiguation of the transcript and each of a plurality of different sequences of words included in the candidate word pool, a plurality of candidate sentences, wherein designating the plurality of candidate sentences further includes determining a confidence score for the transcript and each of the plurality of candidate sentences; determining, by the software code executed by the processing hardware using at least one of an entity identified in the transcript or a history of the user, at least one of an intent of the user or a context of the speech by the user; performing a ranking of the plurality of candidate sentences and the transcript, by the software code executed by the processing hardware, based on the at least one of the intent of the user or the context of the speech by the user; and identifying, by the software code executed by the processing hardware based on the ranking, one of the plurality of candidate sentences or the transcript as a best transcription of the speech by the user.