Speech Recognition Corpus Error Detection Using Perplexity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting transcription errors in speech recognition corpora are inadequate, as they either focus on speech recognition results alone or ignore text characteristics, leading to unreliable learning outcomes for speech recognition models.
Innovation Solution
A method and device that utilize a speech recognition model and a language model to extract performance evaluation indices and perplexities, automatically detecting transcription errors by setting primary and secondary error candidates based on predefined reference values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing error detection methods are used, then the process is simple, but the detection accuracy is insufficient because they only consider speech recognition results without considering text characteristics
Solution Approach 1:
The patent combines speech recognition results with language model predictions to create a comprehensive error detection system. By merging the speech recognition output with language model probability assessments, the system achieves higher detection accuracy while maintaining reasonable complexity through integrated processing.
Solution Approach 2:
The language model acts as an intermediary between the speech recognition system and the error detection process. It provides probability scores that mediate the comparison between transcribed text and expected language patterns, enabling accurate error detection without directly analyzing raw speech signals.
2Productivity
If manual inspection of large corpora is performed, then detection accuracy is high, but time consumption increases significantly
Solution Approach 1:
The system performs self-service by automatically detecting transcription errors using the integrated speech recognition and language model approach. This eliminates the need for manual inspection while maintaining high detection accuracy, thereby significantly reducing time consumption and improving productivity.
Solution Approach 2:
The patent replaces manual mechanical inspection with an automated computational system. By substituting human reviewers with an automated pipeline that combines speech recognition and language model analysis, the system achieves scalable error detection without time loss.
3Reliability
If transcription errors are not detected, then the corpus can be used as-is, but the reliability of speech recognition model learning is compromised
Solution Approach 1:
The system performs preliminary action by automatically detecting and identifying transcription errors before the corpus is used for training. This preliminary error detection ensures that only reliable data is used for learning, thereby improving reliability without requiring extensive manual corpus preparation.
Solution Approach 2:
The language model provides feedback on the quality of transcriptions by comparing predicted probabilities against actual transcribed text. This feedback mechanism enables automatic identification of errors, ensuring reliable learning data while simplifying corpus preparation efforts.
Data Source
AI summary
Provided are a method and device for detecting transcription error in a speech recognition corpus. The method of detecting a transcription error in speech recognition corpus includes following steps: (a) receiving the speech recognition corpus including a speech file and a text label for the speech file; (b) performing speech recognition on the speech file of the speech recognition corpus using a speech recognition model and converting the speech recognition result into text; (c) extracting a performance evaluation index of the speech recognition model; (d) extracting a PPL(s2) for the text label and a PPL(s1) for the text using a language model; and (e) detecting a transcription error in text label of the speech recognition corpus using the extracted performance evaluation index and the PPL(s2) and PPL(s1).