ASR Personalization Using Anti-Context Examples Against Overlearning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition (ASR) systems often generate inaccurate transcriptions due to misrecognizing spoken phrases, leading to overlearning and loss of ability to correctly transcribe common phrases, especially when correcting unique or uncommon words.
Innovation Solution
Utilize anti-context examples containing misrecognized phrases paired with text-to-speech (TTS) audio data to prevent overlearning by training the speech recognition model to recognize both intended and unintended phrases accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the speech recognition model is updated using only user-corrected text and audio data, then the model learns to recognize unique or uncommon words accurately, but the model loses the ability to correctly transcribe common phrases due to overlearning
Solution Approach 1:
The patent applies preliminary anti-action by introducing anti-context examples that preemptively prevent the model from overlearning. These examples contain misrecognized common phrases paired with TTS audio, training the model to recognize that certain phonetic patterns correspond to common words rather than unique words, thereby preventing the model from incorrectly adapting to uncommon word patterns during personalization
2Adaptability or versatility
If the speech recognition model is personalized based on user corrections, then the model adapts to the user's speech patterns and uncommon words, but the model generates more misrecognitions for common phrases
Solution Approach 1:
The patent introduces anti-context examples as an intermediary training component between the user's unique speech patterns and the model's transcription output. These intermediary examples serve as a buffer that teaches the model to distinguish between user-specific uncommon words and standard common phrases, allowing the model to adapt to user patterns without sacrificing accuracy on common vocabulary
3Measurement precision
If the speech recognition model is trained on user-corrected data only, then the model improves recognition for the user's uncommon words, but the model fails to maintain general speech recognition capabilities
Solution Approach 1:
The patent applies parameter changes by modifying the training data composition to include both user-corrected examples and anti-context examples with misrecognized common phrases. This changes the statistical parameters of the training distribution, ensuring the model learns to recognize both uncommon user words and common phrases by exposing it to a balanced dataset that preserves general recognition capabilities while accommodating user-specific patterns
Data Source
AI summary
A method (400) for using anti-context examples for personalizing a speech recognition model (132) includes receiving audio data (104) corresponding to an utterance (102) spoken by a user (10), and processing, using the speech recognition model, the audio data to generate a transcription (106) of the utterance. The transcription including a misrecognized phrase (144) that was misrecognized in the transcription by the speech recognition model. The method also includes receiving user-corrected text (141) including a corrected phrase (146) that replaces the misrecognized phrase that was misrecognized in the transcription. Based on the misrecognized phrase, the method includes generating an anti-context example (305) including anti-context text (310) containing the misrecognized phrase paired with text-to-speech (TTS) audio data (315) corresponding to a synthesized speech representation of the anti-context text. The method also includes personalizing the speech recognition model based on the anti-context example.