Ensemble ASR Accuracy via Fixed-Width Error Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Individual automatic speech recognition (ASR) engines vary in performance under different conditions, such as languages, accents, dialects, voices, and audio characteristics, leading to inconsistent transcription accuracy.
Innovation Solution
A method is introduced to analyze transcriptions by storing per-character differences between two transcriptions in a matrix and encoding these differences into a fixed-width vector, which represents the pairwise error rate between the transcriptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple ASR engines are used to generate transcriptions under varying conditions, then coverage of different languages, accents, and dialects is improved, but consistency and reliability of transcription accuracy deteriorate
Solution Approach 1:
The patent combines multiple ASR engine transcriptions into an ensemble model that aggregates results from diverse engines. By merging the strengths of individual engines (each optimized for specific languages, accents, or dialects) into a unified ensemble system, the solution achieves both broad adaptability across different speech conditions and consistent reliable accuracy through collective decision-making of multiple engines.
2Measurement precision
If per-character differences between transcriptions are analyzed in detail, then measurement precision of error rates is improved, but device complexity and computational requirements increase
Solution Approach 1:
The patent extracts only the essential per-character difference information between transcriptions and stores it in fixed-width matrices. By taking out and isolating only the critical error characteristics (character-level differences) while discarding redundant information, the system achieves high measurement precision for pairwise error rates without the computational burden of analyzing complete transcription sequences or unnecessary details.
Solution Approach 2:
The patent transforms variable-length transcription difference data into fixed-width matrix representations. By changing the parameter structure from unbounded text comparisons to standardized fixed-dimensional matrices, the system enables efficient computational processing while maintaining precise measurement of pairwise error rates. This parameter transformation allows consistent mathematical operations across all transcription pairs regardless of their original length variations.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
One embodiment of the present invention sets forth a technique for analyzing transcriptions of a recording. The technique includes storing per-character differences between a first set of characters from a first transcription of the recording and a second set of characters from a second transcription of the recording in a matrix with a fixed width. The technique also includes encoding the per-character differences in the matrix into a vector of the fixed width. The technique further includes outputting the vector as a representation of a pairwise error rate between the first transcription and the second transcription.