Acoustic Model Target Label Segmentation for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
End-to-end speech recognition techniques using deep neural networks face accuracy issues due to the lack of reflection of structural characteristics of recognition target languages, such as Korean, which combine initial, medial, and final consonants in syllable formation.
Innovation Solution
A method is developed to construct and learn a deep neural network-based acoustic model that reflects the structural characteristics of the recognition target language by generating target labels that distinguish consonants based on their positions, using a multi-level target label approach and adjusting the computation process to improve learning speed and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If end-to-end speech recognition is implemented using deep neural networks without analyzing pronunciation, then the recognition process is simplified and processing speed is improved, but the accuracy deteriorates because structural characteristics of the target language are not reflected
Solution Approach 1:
The patent segments the target label into multiple levels (first level: letter units, second level: syllable units, third level: word units) to capture different granularities of language structure. This segmentation allows the acoustic model to learn both fine-grained phonetic information and coarser linguistic patterns, resolving the contradiction between simplified processing and accurate recognition.
Solution Approach 2:
The patent introduces a new dimension of target label construction by incorporating multiple levels of linguistic organization (letter, syllable, word) rather than using a single level. This dimensional expansion enables the model to capture structural characteristics of the target language while maintaining end-to-end processing efficiency.
2Device complexity
If target labels are constructed without distinguishing consonant positions, then the label construction process is simplified, but the accuracy deteriorates because language structural characteristics are not reflected
Solution Approach 1:
The patent segments consonants into different types based on their position in syllables (initial, medial, final) and creates separate target labels for each type. This segmentation adds linguistic information without significantly increasing overall system complexity, as the labeling process follows systematic rules.
Solution Approach 2:
The patent applies local quality by differentiating consonants based on their specific positions and roles within syllables. Each consonant type (initial, medial, final) receives distinct labeling treatment to reflect its unique linguistic characteristics, thereby improving recognition accuracy while maintaining manageable complexity.
3Measurement precision
If multi-level target labels are constructed to reflect language structure, then speech recognition accuracy is improved, but the learning time increases
Solution Approach 1:
The patent performs preliminary action by pre-defining the multi-level target label structure and consonant type classifications before training begins. This pre-prepared labeling schema guides the learning process and allows the model to efficiently learn linguistic patterns without requiring complex real-time analysis during training.
Solution Approach 2:
The patent utilizes parameter changes by adjusting the depth and granularity of target label construction based on the specific requirements of different languages and recognition tasks. This flexibility allows optimization of learning time while maintaining accuracy by adapting the multi-level label structure to the specific linguistic characteristics being modeled.
Data Source
AI summary
Provided is a speech recognition method for a recognition target language. According to an embodiment of the inventive concept, a speech recognition method for a recognition target language performed by a speech recognition apparatus includes obtaining an original learning data set for the recognition target language, constructing a target label by dividing the text information included in each piece of original learning data in letter units, and building an acoustic model based on a deep neural network by learning the learning speech data included in the each piece of original learning data and the target label corresponding to the learning speech data.