Context-Sensitive Speech Recognition via Triphone Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems are not sufficiently reliable, especially in noisy environments, leading to operational deterioration and user dissatisfaction due to insufficient recognition and malfunctions.
Innovation Solution
The method generates phonetic representations of speech signals using context-sensitive modeling, including triphone models and polyphone models, to create variants of the initial recognition result, which are then scored and selected for further processing, ensuring better phonetic approximation and matching actual contexts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If context dependent models (triphones) are used during phonetic recognition stage, then phonetic approximation quality is improved, but computational complexity and processing burden increase due to huge number of combinatorial possibilities
Solution Approach 1:
The patent segments the phoneme recognition process into multiple stages: first generating an initial phonetic sequence, then separately generating context information, and finally combining them. This segmentation allows the system to handle complex triphone modeling without being overwhelmed by the combinatorial explosion of all possible phoneme-context combinations, as each component is processed independently and then integrated.
Solution Approach 2:
The patent performs preliminary generation of phonetic sequences and context information before final combination. By pre-generating the initial phonetic sequence from the speech signal and separately pre-generating context information from language models, the system prepares components that can be efficiently combined later, reducing the real-time computational burden during actual recognition.
2Reliability
If vocabulary size is increased to improve recognition coverage, then recognition accuracy is improved, but storage capacity requirements increase
Solution Approach 1:
The patent divides the large vocabulary handling into manageable segments by processing speech signals through multiple recognition stages. Instead of loading and processing all vocabulary entries simultaneously, the system segments the recognition process into phoneme-level processing, context generation, and sequence combination, allowing efficient handling of large vocabularies with reduced storage demands at any single processing stage.
3Reliability
If multi-pass recognition process is used to improve reliability, then recognition accuracy is improved, but processing time increases
Solution Approach 1:
The patent performs preliminary generation of phonetic sequences and context information in advance, before the final combination stage. By pre-computing the initial phonetic sequence from the speech signal and pre-generating context information from language models, the system reduces the computational burden during the actual multi-pass recognition process, thereby minimizing additional processing time while maintaining improved reliability.
4Measurement precision
If context information is enforced to match between models and final sequence, then phonetic accuracy is improved, but errors are introduced due to combinatorial possibilities
Solution Approach 1:
Instead of forcing context information to match between models and final sequence (which causes errors), the patent inverts the approach: it generates the initial phonetic sequence first, then generates context information separately, and finally combines them. This inversion allows the system to accommodate natural variations in context without forcing rigid matching, thereby maintaining phonetic accuracy while avoiding the introduction of errors from combinatorial constraints.
Data Source
Figure 1
Figure 2
AI summary
The present invention relates to a method for speech recognition and a method for voice enrolment, comprising the steps of detecting and digitizing a speech signal representing a verbal utterance; generating a phonetic representation of the speech signal as a first recognition result; generating variants of the phonetic representation; and selecting at least one of the variants of the phonetic representation as a second recognition result.