Automatic Language Model Selection for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech to text (STT) systems face a balance between out-of-vocabulary (OOV) error rate and system performance, as using a large vocabulary reduces OOV but increases error rates, while a small vocabulary improves performance but increases OOV errors.
Innovation Solution
The system generates a phonetic lattice from audio input, analyzes it using a first language model, and selects a second language model based on certainty values and match values to generate a final transcription, combining sub-words into words and using a combined index to improve language model selection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large vocabulary is used in STT systems, then the out-of-vocabulary error rate is reduced, but the system performance deteriorates and word error rate increases
Solution Approach 1:
The patent segments the vocabulary into multiple domains or categories, allowing the system to select only the relevant vocabulary subset for each specific speech input. This segmentation resolves the contradiction by maintaining a large overall vocabulary for coverage while using small domain-specific vocabularies for accurate recognition, thus reducing both OOV and word errors simultaneously.
Solution Approach 2:
The patent implements dynamic vocabulary selection where the system adaptively chooses the appropriate vocabulary subset based on the characteristics of the input speech. This dynamic approach allows the system to optimize between vocabulary coverage and recognition accuracy in real-time, resolving the static contradiction between large and small vocabulary usage.
2Manufacturing precision
If a small vocabulary is used in STT systems, then system performance improves, but the out-of-vocabulary error rate increases
Solution Approach 1:
The patent creates a universal STT system framework that can function with different vocabulary subsets. The system maintains multiple vocabulary databases and selectively applies the appropriate one based on input characteristics, thus achieving both high performance with small vocabularies and low OOV rates through multi-functionality.
Solution Approach 2:
The patent performs preliminary classification of the input speech to determine which vocabulary subset is most appropriate before proceeding with recognition. This preliminary action allows the system to prepare the optimal vocabulary configuration in advance, ensuring both high performance and low OOV errors without requiring a full large vocabulary during processing.
Data Source
AI summary
A system and method for generating a transcript of an audio input. An embodiment of a system and method may include generating a phonetic lattice by decoding the audio input and producing a transcription based on the phonetic lattice and based on a first language model. A transcription may be analyzed to produce analysis results. Analysis results may be used to select from a plurality of language models, one language model and the selected language model may be used to generate a transcript of the audio input.