Dynamic Language Model Biasing for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated Speech Recognition (ASR) systems face challenges in accurately transcribing user utterances, especially in contexts where general language models do not fit well, leading to reduced recognition quality due to high latency and inability to predict expected phrases with sufficient probability.
Innovation Solution
A multi-pass speech recognition system that analyzes a recognition lattice from a first pass to determine if a context-specific language model should be used in a second pass, biasing the speech recognition towards the specific context by combining a general language model with a context language model, thereby improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a general language model is used for speech recognition, then the system can handle diverse utterances, but recognition accuracy deteriorates in specific contexts where the general model does not fit well
Solution Approach 1:
The system dynamically switches between general language models and context-specific language models based on the recognition lattice analysis. The biasing mechanism is adjusted in real-time during multi-pass recognition, making the system adaptable rather than static. This resolves the contradiction by allowing the system to be general when needed but specialized when accuracy requires it.
Solution Approach 2:
The patent applies different language model qualities to different parts of the recognition process. Context-specific language models with higher precision are applied locally to specific contexts identified in the recognition lattice, while general models handle other cases. This local application of specialized models improves accuracy where needed without sacrificing overall versatility.
2Measurement precision
If context-specific language models are used to improve recognition accuracy, then transcription quality improves, but system complexity increases due to multiple passes and model selection
Solution Approach 1:
The system performs preliminary analysis of the recognition lattice in the first pass to identify contexts where context-specific language models should be applied. This preliminary action allows the second pass to focus computational resources only on problematic areas, improving transcription quality without requiring complete re-processing of all utterances through complex models.
Solution Approach 2:
The patent applies context-specific language models partially - only to the extent necessary for improving accuracy in identified contexts. The multi-pass system performs speech recognition twice, but the second pass is biased only towards contexts where the general model failed, rather than applying complex models universally. This selective application reduces overall system complexity while maintaining improved transcription quality where needed.
3Measurement precision
If multi-pass speech recognition is performed with context biasing, then recognition accuracy improves for specific contexts, but processing time increases due to multiple recognition passes
Solution Approach 1:
The first pass of speech recognition performs preliminary transcription and generates the recognition lattice, identifying contexts where improvement is needed. This preliminary action allows the system to prepare context information in advance, so the second pass can be biased efficiently towards specific contexts rather than performing exhaustive search, reducing the time penalty of multi-pass processing.
Solution Approach 2:
The second pass performs speech recognition with context biasing only for identified problematic contexts, rather than re-processing all utterances with full complexity. This partial action approach improves recognition accuracy for specific contexts while minimizing the additional processing time required, as the biased recognition focuses computational effort only where needed.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for speech recognition. In one aspect, a method comprises receiving audio data encoding one or more utterances; performing a first speech recognition on the audio data; identifying a context based on the first speech recognition; performing a second speech recognition on the audio data that is biased towards the context; and providing an output of the second speech recognition.