Speech Model Refinement via Transcription Error Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition model training is hindered by transcription errors in training data, which degrade recognition performance, especially as model accuracy improves, and manual error correction is expensive and time-consuming.
Innovation Solution
A reliable transcription error-checking algorithm using word confidence scores and word duration probabilities is employed to detect and correct errors, allowing for iterative model retraining to improve speech recognition quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual transcription error correction is performed, then transcription accuracy is improved, but processing time and cost increase significantly
Solution Approach 1:
The system uses the trained speech recognition model itself to detect transcription errors in the training data. The model automatically identifies potential errors by analyzing acoustic features and comparing them against learned patterns, eliminating the need for external manual review and enabling self-correction of training data
Solution Approach 2:
The patent replaces manual mechanical review processes with automated computational analysis. The system uses algorithmic error detection based on acoustic modeling, likelihood ratios, and confidence scoring to substitute human transcribers, dramatically reducing processing time while maintaining high accuracy
2Measurement precision
If larger training datasets are used, then model accuracy is improved, but the impact of transcription errors increases
Solution Approach 1:
The system performs preliminary error detection and filtering on the training dataset before model training begins. By identifying and removing or correcting erroneous transcriptions in advance, the system ensures that only high-quality data is used for training, preventing error propagation even in large datasets
Solution Approach 2:
The patent implements a feedback loop where the trained model continuously evaluates the quality of training data. Detected errors are corrected and fed back into the training set, allowing iterative improvement of both data quality and model performance, thereby reducing the cumulative impact of errors in large datasets
3Reliability
If conventional error processing methods are used, then some errors are corrected, but the process is expensive and unreliable
Solution Approach 1:
The system changes the parameters used for error detection from simple keyword matching or basic quality metrics to sophisticated acoustic feature analysis. By utilizing phone-level alignment, duration probabilities, and confidence scores derived from the speech model, the system achieves reliable error detection with a unified processing framework rather than multiple complex separate processes
Data Source
AI summary
Reliable transcription error-checking algorithm that uses a word confidence score and a word duration probability to detect transcription errors for improved results through the automatic detection of transcription errors in a corpus. The transcription error-checking algorithm is combined model training so as to use a current model to detect transcription errors, remove utterances which contain incorrect transcription (or manually fix the found errors), and retrain the model. This process can be repeated for several iterations to obtain an improved speech recognition model. The speech model is employed to achieve speech-transcription alignment to obtain a word boundary. Speech recognizer is then utilized to generate a word-lattice. Using the word boundary and word lattice, error detection is computed using a word confidence score and a word duration probability.