Contextual Spelling Correction for ASR With Large Context Lists
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic speech recognition (ASR) systems face challenges in accurately transcribing context-specific words and specialized terminology due to high word error rates and latency issues, especially with large context lists, and prior solutions like contextual language models and biased encoders suffer from performance degradation.
Innovation Solution
A contextual spelling correction (CSC) model is applied on top of the ASR model, using a shared context encoder and filtering mechanism to handle large context lists, allowing run-time customization without altering the original ASR model, and incorporating context information for spelling correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If contextual language model or biased encoder is used to improve transcription accuracy for context-specific words, then word error rate decreases, but latency increases and performance degrades for large context lists
Solution Approach 1:
The system segments the context list into manageable units by maintaining a fixed-size buffer of recent context-specific words. Instead of processing the entire context list at once, the CSC model processes only the most relevant recent contexts, reducing computational overhead and latency while maintaining transcription accuracy for context-specific terms.
Solution Approach 2:
The system extracts only the necessary context information needed for spelling correction. The CSC model is trained to identify and correct errors in context-specific words without requiring the entire context list to be processed, thereby reducing the amount of data that needs to be handled and improving response time.
2Measurement precision
If contextual language model or biased encoder is used to improve transcription accuracy for specialized terminology, then word error rate decreases, but system complexity increases
Solution Approach 1:
The CSC model acts as an intermediary component that sits between the standard ASR model and the final output. It receives ASR hypotheses and context lists, performs targeted spelling correction for context-specific words, and returns corrected results. This modular approach adds functionality without fundamentally altering the core ASR model architecture.
Solution Approach 2:
The CSC model is designed to handle multiple types of context-specific vocabulary (contact names, specialized terminology, domain-specific terms) through a single unified model. This multi-functional design avoids the need for separate specialized models for different types of context, thereby reducing overall system complexity.
3Measurement precision
If contextual spelling correction is applied to handle large context lists, then accuracy for out-of-vocabulary terms improves, but computational resources increase
Solution Approach 1:
The system applies partial action by processing only the top-K most probable ASR hypotheses through the CSC model rather than all possible hypotheses. This selective processing reduces computational resources while still achieving high accuracy for out-of-vocabulary terms in the most likely transcriptions.
Solution Approach 2:
The CSC model is pre-trained on context-specific vocabulary and patterns before deployment. This preliminary training allows the model to efficiently recognize and correct context-specific terms during runtime without requiring extensive computational resources for real-time analysis of every possible word combination.
Data Source
AI summary
Novel solutions for speech recognition provide contextual spelling correction (CSC) for automatic speech recognition (ASR). Disclosed examples include receiving an audio stream; performing an ASR process on the audio stream to produce an ASR hypothesis; receiving a context list; and, based on at least the ASR hypothesis and the context list, performing spelling correction to produce an output text sequence. A contextual spelling correction (CSC) model is used on top of an ASR model, precluding the need for changing the original ASR model. This permits run-time user customization based on contextual data, even for large-size context lists. Some examples include filtering ASR hypotheses for the audio stream and, based on at least the ASR hypotheses filtering, determining whether to trigger spelling correction for the ASR hypothesis. Some examples include generating text to speech (TTS) audio using preprocessed transcriptions with context phrases to train the CSC model.