Large Vocabulary Speech Recognition Using Acoustic Invariant Structure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for continuous speech recognition using acoustic invariant structure are not effective for large vocabulary tasks due to the lack of suitable decoding algorithms, and previous approaches have not significantly improved recognition rates.
Innovation Solution
A large vocabulary continuous speech recognition method that calculates a structure score for each hypothesis by multiplying the likelihood of inter-distribution distance of phoneme pairs by weighting factors, normalizes the likelihood, and ranks hypotheses based on both speech recognition scores and structure scores, with pair-by-pair weightings learned using training data to prioritize phoneme relationships.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If acoustic invariant structure is used for continuous speech recognition, then robustness against non-linguistic factors is improved, but recognition rate deteriorates due to lack of suitable decoding algorithms
Solution Approach 1:
The patent segments the hypothesis evaluation into two independent parts: traditional speech recognition scoring and acoustic invariant structure scoring. Each component is calculated separately and then combined, allowing both robustness and recognition rate to be optimized independently before integration.
Solution Approach 2:
The patent introduces a new scoring parameter (acoustic invariant structure score) that complements the traditional speech recognition score. By changing the evaluation parameters from a single metric to multiple metrics, the system achieves both robustness and high recognition rate through multi-dimensional hypothesis evaluation.
2Reliability
If phoneme alignment is performed for each hypothesis, then acoustic invariant structure can be extracted, but computational complexity increases
Solution Approach 1:
The patent performs phoneme alignment and extracts acoustic invariant structure as preliminary actions during the N-best hypothesis generation phase, rather than as a separate post-processing step. This integration reduces overall computational complexity by reusing intermediate results across multiple hypotheses.
Solution Approach 2:
The patent creates a universal phoneme alignment process that serves multiple functions: it enables acoustic invariant structure extraction, provides basis for structure score calculation, and supports hypothesis ranking. This multi-functional approach eliminates redundant computations.
3Measurement precision
If structure score is calculated for all phoneme pairs, then accurate hypothesis ranking is achieved, but processing time increases
Solution Approach 1:
The patent applies local quality by calculating structure scores selectively for critical phoneme pairs that have significant impact on hypothesis discrimination, rather than uniformly for all phoneme pairs. This targeted approach maintains ranking accuracy while reducing processing time.
Solution Approach 2:
The patent reuses phoneme alignment results and inter-distribution distance calculations across multiple hypotheses, copying intermediate results rather than recalculating them independently for each hypothesis. This significantly reduces redundant computations and processing time.
Data Source
AI summary
System and method for performing speech recognition using acoustic invariant structure for large vocabulary continuous speech. An information processing device receives sound as input and performs speech recognition. The information processing device includes: a speech recognition processing unit for outputting a speech recognition score, a structure score calculation unit for calculation of a structure score that is a score that, with respect for each hypothesis concerning all phoneme pairs comprising the hypothesis, is found by applying phoneme pair-by-pair weighting to phoneme pair inter-distribution distance likelihood and then performing summation, and a ranking unit for ranking the multiple hypotheses based on a sum value of speech recognition score and structure score.