Speech Recognition Rescoring Using Prosodic Likelihood
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition techniques for languages using prosody, such as tones, face inaccuracies when tone recognition is applied to consonants, leading to lower speech recognition accuracy.
Innovation Solution
A speech recognition apparatus and method that performs word search based on acoustic distance and phonetic models, outputs word hypotheses and prosodic labels, and corrects scores using prosodic recognition to accurately identify tone intervals within vowel segments, thereby improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If tone recognition is applied to all syllable portions including consonants, then the speech recognition system can process more data, but the recognition accuracy decreases because consonants do not have voice tones
Solution Approach 1:
The syllable is divided into two parts: the first half portion (which may include consonants) and the second half portion (which contains the vowel with tone). Tone recognition is applied only to the second half portion where tones actually occur, while the first half portion is processed separately. This segmentation resolves the contradiction by limiting tone recognition to appropriate regions, maintaining both processing coverage and accuracy.
Solution Approach 2:
Different parts of the syllable are treated with different recognition strategies. The first half portion (consonant region) is processed without tone association, while the second half portion (vowel region) is processed with tone recognition. This local differentiation ensures that tone recognition accuracy is maintained by applying the technique only where it is valid, while still processing the entire syllable structure.
2Productivity
If tone recognition is applied to consonants, then more syllables can be processed, but false tone recognition occurs leading to lower speech recognition accuracy
Solution Approach 1:
The syllable is segmented into a first half portion and a second half portion, with tone recognition restricted to the second half portion. This ensures that only syllable portions that actually contain tones are processed with tone recognition, preventing false recognition while maintaining processing of all syllables through the combined approach of both portions.
Solution Approach 2:
The tone recognition function is extracted and applied only to the second half portion of the syllable, separating the tone recognition operation from the consonant processing. This extraction ensures that tone recognition does not incorrectly apply to consonants, maintaining reliability while still processing the complete syllable structure through the first half portion handling.
Data Source
AI summary
A distance calculation unit (16) obtains the acoustic distance between the feature amount of input speech and each phonetic model. A word search unit (17) performs a word search based on the acoustic distance and a language model including the phoneme and prosodic label of a word, and outputs a word hypothesis and a first score representing the likelihood of the word hypothesis. The word search unit (17) also outputs a vowel interval and its tone label in the input speech, when assuming that the recognition result of the input speech is the word hypothesis. A tone recognition unit (21) outputs a second score representing the likelihood of the tone label output from the word search unit (17) based on a feature amount corresponding to the vowel interval output from the word search unit (17). A rescore unit (22) corrects the first score of the word hypothesis output from the word search unit (17) using the second score output from the tone recognition unit (21). This allows to raise the speech recognition accuracy for tone speech.