Vocal Tract Length Normalization for Low Latency Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face challenges in normalizing vocal tract lengths across different speakers, leading to significant acoustic differences and latency issues, even when using Vocal Tract Length Normalization (VTLN) techniques that require substantial speech data for effective speaker normalization.
Innovation Solution
A method and apparatus for generating a Vocal Tract Length Normalized acoustic model from training data, estimating normalization factors based on recognition inputs, and applying these factors to improve speech recognition accuracy with reduced latency by using short speech segments for warp estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Vocal Tract Length Normalization is applied using traditional methods with multiple minutes of speech per speaker, then speaker normalization accuracy is improved, but recognition latency increases significantly
Solution Approach 1:
The patent pre-calculates and stores VTLN warping factors during the training phase for each speaker. When recognition is needed, these pre-computed factors are directly applied without requiring real-time calculation, thus maintaining high normalization accuracy while eliminating the time delay that would occur during actual speech recognition
Solution Approach 2:
The patent uses a fixed set of discrete VTLN warping factors (typically 5-9 factors ranging from 0.8 to 1.2) instead of continuously adjusting factors based on entire speech segments. This partial approach using selected discrete factors achieves sufficient normalization accuracy while dramatically reducing computational complexity and latency compared to traditional methods that analyze minutes of speech
2Productivity
If Vocal Tract Length Normalization is applied in real-time, then recognition speed is improved, but normalization accuracy deteriorates due to insufficient speech data
Solution Approach 1:
VTLN warping factors are computed during the training phase when sufficient speech data is available, and these pre-computed factors are stored for later use. This allows real-time recognition to use accurate normalization factors without needing to re-analyze speech data, thus maintaining both speed and accuracy
Solution Approach 2:
The patent creates a copy of the VTLN warping factors during training that can be applied during recognition without modifying the original training data. This copied normalization model enables real-time processing with the same accuracy as if full speech analysis were performed, since the normalization characteristics are captured and reused
3Reliability
If traditional VTLN methods are used requiring minutes of speech per speaker, then normalization effectiveness is improved, but system complexity and data requirements increase
Solution Approach 1:
The patent applies a fixed set of discrete VTLN warping factors (5-9 factors) rather than continuously varying factors based on extensive speech analysis. This partial approach using predetermined factors achieves sufficient normalization effectiveness while dramatically simplifying the system by eliminating the need for complex real-time speech analysis and storage of large speech datasets
Solution Approach 2:
The patent transforms the continuous VTLN warping factor into a discrete set of predefined values (typically 0.8, 0.85, 0.9, 0.95, 1.0, 1.05, 1.1, 1.2). This parameter discretization simplifies the normalization process by allowing selection from fixed options rather than computing continuous values, reducing both computational complexity and data requirements while maintaining normalization effectiveness
Data Source
AI summary
A method and system for training an automatic speech recognition system are provided. The method includes separating training data into speaker specific segments, and for each speaker specific segment, performing the following acts: generating spectral data, selecting a first warping factor and warping the spectral data, and comparing the warped spectral data with a speech model. The method also includes iteratively performing the steps of selecting another warping factor and generating another warped spectral data, comparing the other warped spectral data with the speech model, and if the other warping factor produces a closer match to the speech model, saving the other warping factor as the best warping factor for the speaker specific segment. The system includes modules configured to control a processor in the system to perform the steps of the method.