Language Model Building for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition methods face challenges in accurately recognizing nonstandard pronunciations, dialects, and mispronunciations, leading to increased error rates and loss of speech information, particularly when dealing with different languages and pronunciation habits.
Innovation Solution
A method for building a language model that includes receiving candidate sentences, obtaining phonetic spellings and probabilities, and using an acoustic model and syllable acoustic lexicon to match phonetic transcriptions with text sequences, thereby selecting the most probable text sequence as the recognition result.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If phonetic transcriptions marked in the dictionary are used for training the acoustic model, then the acoustic model can be trained with standard pronunciations, but nonstandard pronunciations will increase fuzziness and mismatch with the phonetic transcriptions, leading to increased error rates
Solution Approach 1:
The patent segments the phonetic transcription process into multiple stages: first obtaining phonetic transcriptions from the acoustic model, then converting them to phonetic spellings through the acoustic lexicon, and finally matching with the language model. This segmentation allows each component to handle specific aspects of pronunciation variation independently, improving overall accuracy for both standard and nonstandard pronunciations
Solution Approach 2:
The patent introduces phonetic spellings as an intermediary layer between phonetic transcriptions and vocabularies. The phonetic spellings serve as a flexible intermediate representation that can accommodate various pronunciation variations before being mapped to final vocabulary items, reducing the direct mismatch between standard transcriptions and nonstandard pronunciations
2Productivity
If phonetic transcriptions are converted into vocabularies by the acoustic lexicon, then vocabularies can be obtained for further processing, but speech information such as accent locations is lost, influencing accuracy in intention recognition and increasing error rates in semanteme recognition
Solution Approach 1:
The patent performs preliminary matching of phonetic spellings with the language model before final vocabulary conversion. By conducting the matching operation in advance with the flexible phonetic spelling representation, the system captures pronunciation variations and accent information early in the processing chain, preserving this information for subsequent semanteme recognition and intention analysis
Solution Approach 2:
The patent incorporates feedback mechanisms where the language model provides probability information about candidate vocabularies based on phonetic spellings. This feedback loop allows the system to refine its selections and preserve important speech information by considering contextual likelihoods, thereby improving semanteme recognition accuracy while maintaining processing efficiency
Data Source
AI summary
A method for building a language model, a speech recognition method and an electronic apparatus are provided. The speech recognition method includes the following steps. Phonetic transcriptions of a speech signal are obtained from an acoustic model. Phonetic spellings matching the phonetic transcriptions are obtained according to the phonetic transcriptions and a syllable acoustic lexicon. According to the phonetic spellings, a plurality of text sequences and a plurality of text sequence probabilities are obtained from a language model. Each phonetic spelling is matched to a candidate sentence table; a word probability of each phonetic spelling matching a word in a sentence of the sentence table are obtained; and the word probabilities of the phonetic spellings are calculated so as to obtain the text sequence probabilities. The text sequence corresponding to a largest one of the sequence probabilities is selected as a recognition result of the speech signal.