Acoustic Confusability Measures for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems rely on lexical perplexity as an objective function for training language models, which is a poor indicator of model quality and can lead to increased errors, necessitating a more effective measure for assessing language model quality.
Innovation Solution
The development of empirically derived acoustic confusability measures that are computed efficiently and iteratively, allowing for the determination of acoustic confusability between textual phrases without relying on internal speech recognition models, and the use of these measures to improve speech recognition accuracy by identifying confusable phrases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If lexical perplexity is used as the objective function for training language models, then the training process is simplified and computationally efficient, but the model quality assessment becomes inaccurate leading to increased speech recognition errors
Solution Approach 1:
The patent introduces acoustic confusability measures as an intermediary metric between the training process and model quality assessment. Instead of directly using lexical perplexity which fails to capture acoustic similarity, the system computes acoustic confusability between pairs of words using acoustic models and pronunciation information. This intermediary measure provides a more accurate indication of model quality while maintaining computational feasibility through efficient algorithms.
Solution Approach 2:
The patent replaces the traditional lexical perplexity metric with an acoustic-based confusability measure. This substitution transitions from a purely lexical/statistical approach to one that incorporates acoustic properties and human perception of speech similarity. The new metric better reflects actual speech recognition difficulty by considering how acoustically similar different words are, rather than just their co-occurrence statistics.
2Measurement precision
If acoustic confusability measures are computed using traditional methods, then the assessment of language model quality becomes more accurate, but the computational complexity and time required increase significantly
Solution Approach 1:
The patent segments the computation of acoustic confusability into manageable components. Instead of computing all possible word pairs simultaneously, the system divides the vocabulary into subsets and computes confusability measures in stages. This segmentation allows for efficient parallel processing and reduces the memory requirements, making the computationally intensive task feasible for practical implementation.
Solution Approach 2:
The patent performs preliminary computations of acoustic confusability measures before the actual speech recognition task. By pre-computing and storing confusability matrices for common word pairs or phoneme sequences, the system avoids repeated heavy computations during real-time operation. This preliminary action caches results that can be quickly retrieved and applied during model training and evaluation.
3Reliability
If acoustic confusability measures are used to identify confusable phrases, then speech recognition accuracy improves, but the system complexity increases due to additional computation and data structures
Solution Approach 1:
The patent applies acoustic confusability measures selectively to only those word pairs or phrases that are most likely to cause confusion, rather than uniformly processing all vocabulary items. By identifying and focusing on high-confusability pairs based on acoustic similarity thresholds or frequency-based prioritization, the system achieves improved recognition accuracy for problematic cases without the full computational overhead of analyzing every possible pair.
Data Source
AI summary
A computer-implemented method includes generating an empirically derived acoustic confusability measure by processing example utterances and iterating from an initial estimate of the acoustic confusability measure to improve the measure. The method can further include using the acoustic confusability measure to selectively limit phrases to make recognizable by a speech recognition application.