Offline Speech Recognition Corpus Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face a trade-off between recognition latency and accuracy, with larger or more complex models providing better accuracy but at the cost of slower processing, while simpler models offer quicker results but with higher error rates, and human transcription of large datasets is laborious and inefficient.
Innovation Solution
An offline speech recognition system with larger or more complex models is used to re-decode utterances from a production system, generating a large number of transcriptions with high accuracy, which are then used to train an updated production speech recognition system, leveraging semi-supervised techniques and distributed computation to create a large training corpus.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If larger or more complex speech recognition models are used, then speech recognition accuracy is improved, but processing speed decreases
Solution Approach 1:
The system segments the speech recognition task into two distinct phases: an offline phase using a complex reference model for high-accuracy transcription, and an online phase using a simplified production model for real-time processing. This segmentation allows each model to be optimized for its specific function, resolving the contradiction between accuracy and speed.
Solution Approach 2:
The production speech recognizer creates a simplified copy or adaptation of the reference model's acoustic and language models. This copy retains the essential recognition capabilities while using fewer computational resources, enabling real-time processing without sacrificing too much accuracy.
2Measurement precision
If human transcription of large datasets is used, then training corpus quality is improved, but time consumption increases
Solution Approach 1:
The system uses the reference model to automatically transcribe speech data for training purposes, making the system self-sufficient in generating training data. This eliminates the need for manual human transcription while maintaining high quality through the reference model's accuracy.
Solution Approach 2:
The reference model performs preliminary transcription of speech data before the simplified production model uses it for training. This preliminary action creates high-quality training corpora in advance, saving significant time compared to manual transcription processes.
Data Source
AI summary
The present disclosure relates to training a speech recognition system. One example method includes receiving a collection of speech data items, wherein each speech data item corresponds to an utterance that was previously submitted for transcription by a production speech recognizer. The production speech recognizer uses initial production speech recognizer components in generating transcriptions of speech data items. A transcription for each speech data item is generated using an offline speech recognizer, and the offline speech recognizer components are configured to improve speech recognition accuracy in comparison with the initial production speech recognizer components. The updated production speech recognizer components are trained for the production speech recognizer using a selected subset of the transcriptions of the speech data items generated by the offline speech recognizer. An updated production speech recognizer component is provided to the production speech recognizer for use in transcribing subsequently received speech data items.