Selective Dynamic Language Models for Speech Transcription Latency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face challenges in efficiently transcribing utterances due to the computational overhead and latency associated with dynamically generating user-specific or utterance-specific class-based language models for every input, which can be unnecessary for utterances that do not include class-based terms.
Innovation Solution
A speech recognition system performs a preliminary decoding using a static class-based language model to determine if an utterance includes class-based terms, and only generates dynamic models for those that require customization, thereby reducing unnecessary computational expense and latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If dynamic class-based language models are generated for every utterance, then transcription accuracy is improved, but computational overhead and latency increase
Solution Approach 1:
The system performs preliminary decoding using a static class-based language model before generating dynamic models. This preliminary action identifies whether an utterance contains class-based terms, allowing the system to prepare only when necessary and reduce overall latency for common cases.
Solution Approach 2:
Instead of generating dynamic models for all utterances, the system applies partial action by generating dynamic models only for a subset of utterances that contain class-based terms. This selective approach reduces computational overhead while maintaining accuracy where needed.
2Measurement precision
If dynamic class-based language models are generated for every utterance, then transcription accuracy is improved, but computational expense increases
Solution Approach 1:
The system generates dynamic class-based language models only for utterances that contain class-based terms, rather than for all utterances. This partial action significantly reduces computational expense and energy consumption while maintaining transcription accuracy for the relevant subset of utterances.
Solution Approach 2:
The system extracts and identifies class-based terms from utterances using preliminary decoding with a static model. By taking out only the necessary computation for utterances containing these terms, the system avoids unnecessary computational expense for the majority of utterances that do not require dynamic modeling.
3Productivity
If static class-based language models are used for all utterances, then computational efficiency is improved, but transcription accuracy deteriorates for utterances with class-based terms
Solution Approach 1:
The system dynamically adjusts the language model used based on the characteristics of each utterance. For utterances containing class-based terms, it switches to dynamic class-based language models, while using static models for other utterances. This dynamic adaptation maintains high computational efficiency overall while improving accuracy where needed.
Solution Approach 2:
The system applies different quality levels of language modeling to different parts of the input space. High-quality dynamic modeling is applied locally to utterances containing class-based terms, while standard static modeling is used for the broader set of utterances, optimizing the balance between efficiency and accuracy.
Data Source
AI summary
A computer-implemented method for transcribing an utterance includes receiving, at a computing system, speech data that characterizes an utterance of a user. A first set of candidate transcriptions of the utterance can be generated using a static class-based language model that includes a plurality of classes that are each populated with class-based terms selected independently of the utterance or the user. The computing system can then determine whether the first set of candidate transcriptions includes class-based terms. Based on whether the first set of candidate transcriptions includes class-based terms, the computing system can determine whether to generate a dynamic class-based language model that includes at least one class that is populated with class-based terms selected based on a context associated with at least one of the utterance and the user.