Contextual Spelling Correction Layer for Large ASR 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 unique spellings and dynamic context lists, leading to high latency and word error rates, particularly when using biased finite-state machines and encoder-based solutions.
Innovation Solution
A contextual spelling correction (CSC) model is applied on top of the ASR model, utilizing a shared context encoder and filtering mechanism to handle large context lists, reducing the need for retraining the original ASR model and enabling run-time customization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a biased finite-state machine or biased encoder is used for contextual spelling correction, then spelling accuracy for context-specific words is improved, but system complexity and training requirements increase
Solution Approach 1:
The system is divided into two independent components: a general ASR model and a contextual spelling correction model. The CSC model operates as a separate post-processing layer that receives ASR hypotheses and context lists, correcting spelling errors without modifying the original ASR model. This segmentation allows each component to be optimized independently, reducing overall system complexity while maintaining high spelling accuracy for context-specific words.
Solution Approach 2:
The contextual spelling correction model serves as an intermediary between the ASR system and the final output. It takes ASR hypotheses and context lists as input, processes them through a shared context encoder and filtering mechanism, and outputs corrected text sequences. This intermediary approach enables spelling correction without requiring changes to the original ASR model, simplifying the system architecture.
2Measurement precision
If a contextual language model with re-scoring is used, then spelling correction for context-specific words is improved, but latency increases for long context lists
Solution Approach 1:
The system performs preliminary filtering of ASR hypotheses before applying full spelling correction. By pre-processing the context list and identifying potential spelling errors in advance, the system can efficiently determine which hypotheses require full CSC processing and which can be handled more lightly, reducing latency for common cases while maintaining accuracy for complex contexts.
Solution Approach 2:
The contextual spelling correction model applies partial correction logic based on the confidence level of ASR hypotheses. For high-confidence hypotheses, the system uses faster, less computationally intensive correction methods, while reserving full re-scoring and context analysis for low-confidence hypotheses that likely contain errors. This partial action approach significantly reduces average latency while maintaining high correction accuracy.
3Measurement precision
If the original ASR model is modified to incorporate contextual spelling correction, then spelling accuracy is improved, but adaptability to dynamic context lists decreases
Solution Approach 1:
The contextual spelling correction model is designed as a dynamic system that adapts to changing context lists in real-time. It uses a shared context encoder that can process variable-length context lists containing contact names, locations, and specialized terminology. The model dynamically adjusts its correction behavior based on the input context, allowing it to handle dynamic and evolving contextual requirements without requiring model retraining or structural changes to the original ASR system.
Solution Approach 2:
The contextual spelling correction model serves multiple functions: it corrects spelling errors for contact names, locations, and specialized terminology; it handles dynamic context lists without requiring model retraining; and it operates independently of the original ASR model. This multi-functionality is achieved through a universal shared context encoder that can process various types of contextual information, making the system highly adaptable to different domains and dynamic requirements.
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.