Single Acoustic Model Multi-Accent Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multi-accent speech recognition in embedded devices is challenging due to the need for multiple acoustic models, which is processor-intensive and costly, especially in environments with limited processing power, and existing solutions require access to acoustic models or symbols recognized by the recognition engine.
Innovation Solution
A speech recognition system that includes an accented phonetic and transformed ID generation unit, which processes metadata and user input speech to provide accent-corrected phonetic data and transformed IDs, allowing for accurate recognition of speech across multiple accents using a single acoustic model, without requiring access to internal acoustic models or symbols.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple accent-specific acoustic models are used to improve recognition accuracy across different accents, then recognition accuracy is improved, but processor usage and resource consumption increase significantly
Solution Approach 1:
The patent merges multiple accent-specific acoustic models into a single unified acoustic model that can handle multiple accents. This is achieved by training the model with diverse accent data and using phonetic transfer pairs to adapt the model to different accents dynamically, thereby reducing the number of separate models needed while maintaining recognition accuracy across different accents
Solution Approach 2:
The patent changes the parameters of a single acoustic model to adapt to different accents by using phonetic transfer pairs. Instead of maintaining separate models with fixed parameters for each accent, the system dynamically adjusts phonetic representations and probability distributions within a single model to match the characteristics of different accents during recognition
2Measurement precision
If multiple accent-specific acoustic models are deployed to handle regional accents, then recognition accuracy is improved, but device complexity and cost increase
Solution Approach 1:
The patent creates a universal acoustic model that serves multiple functions by handling different accents within a single system. The model is designed to be accent-agnostic and can adapt to various accents through phonetic transfer, eliminating the need for separate specialized models for each accent and thereby reducing system complexity
3Measurement precision
If phonetic transfer pairs are constructed to adapt to accented pronunciation, then recognition accuracy for specific accents is improved, but access to acoustic models and internal symbols is required which increases implementation complexity
Solution Approach 1:
The patent introduces phonetic transfer pairs as an intermediary mechanism that bridges the gap between canonical pronunciation and accented pronunciation. These pairs serve as a translation layer that maps standard phonetic representations to accented variants without requiring direct access to the internal symbolic representations of the acoustic model, thereby simplifying implementation
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments are disclosed for recognizing speech in a computing system. An example speech recognition method includes receiving metadata at a generation unit that includes a database of accented substrings, generating, via the generation unit, accent-corrected phonetic data for words included in the metadata, the accent-corrected phonetic data representing different pronunciations of the words included in the metadata based on the accented substrings stored in the database, receiving, at a voice recognition engine, extracted speech data derived from utterances input by a user to the speech recognition system, and receiving, at the voice recognition engine, the accent-corrected phonetic data. The method further includes determining terminal ID(s) identifying recognized utterances in the extracted speech data, generating, accent data identifying accents detected in the recognized utterances, generating recognized speech data based on the one or more terminal IDs and the accent data, and outputting the recognized speech data to the speech-controlled device.