Acoustic Model Training with Shared and Branch Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face challenges in maintaining accuracy due to variations in speech data among native and foreign speakers, as different pronunciation features affect the consistency of acoustic signals, leading to degraded recognition performance.
Innovation Solution
An AI apparatus utilizing a shared network and a branch network to train an acoustic model, where the shared network applies common acoustic information to output hidden representations, and the branch network classifies phonemes specific to each language, allowing for unified phoneme representation across speakers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If speech recognition systems use a single acoustic model for all speakers, then device complexity is reduced, but speech recognition accuracy degrades due to variations in pronunciation among native and foreign speakers
Solution Approach 1:
The acoustic model is segmented into multiple specialized models: a first acoustic model for native speakers and a second acoustic model for foreign speakers. Each model is trained on speech data from its respective speaker group, allowing the system to maintain high recognition accuracy for each group without requiring a single complex universal model. The segmentation enables each model to specialize in the pronunciation characteristics of its target speaker group.
Solution Approach 2:
The system dynamically selects which acoustic model to use based on the detected speaker type. A speaker type detection component identifies whether the current speaker is a native or foreign speaker, and the system switches between the first and second acoustic models accordingly. This dynamic adaptation allows the system to optimize recognition accuracy in real-time without maintaining all possible model variations simultaneously.
2Reliability
If the system trains separate acoustic models for native and foreign speakers, then speech recognition accuracy improves, but training data requirements and processing time increase
Solution Approach 1:
The system performs preliminary classification of speech data into native speaker data and foreign speaker data before training the acoustic models. By pre-separating the training data according to speaker type, the system can train specialized models more efficiently without needing to process all speech data through multiple model training iterations. This preliminary action reduces the overall training time while maintaining the benefits of specialized models.
3Adaptability or versatility
If the acoustic model is designed to handle multiple speaker types, then adaptability improves, but measurement precision of phoneme recognition degrades due to inconsistent acoustic signals
Solution Approach 1:
The system segments the phoneme recognition task by creating separate acoustic models for different speaker groups. The first acoustic model is specialized for native speaker phonemes while the second acoustic model is specialized for foreign speaker phonemes. This segmentation allows each model to achieve high measurement precision for its specific phoneme set without the interference of inconsistent acoustic signals from other speaker groups.
Solution Approach 2:
A speaker type detection component acts as an intermediary between the input speech signal and the acoustic models. This intermediary identifies the speaker type and routes the speech data to the appropriate acoustic model, ensuring that each phoneme recognition task is handled by a model optimized for that speaker group's acoustic characteristics, thereby maintaining high precision.
Data Source
AI summary
Disclosed is an artificial intelligence (AI) apparatus for training an acoustic model, and more particularly, an AI apparatus for training an acoustic model including a shared network and a branch network connected to the shared network using speech data and phonemes corresponding to the speech data.