Improved model adaptation approach for recognition of reduced-frame-rate continuous speech

Area of Science:

• Computational Linguistics
• Signal Processing
• Machine Learning

Background:

• Distributed speech recognition systems rely on front-end devices (smartphones, PDAs) for speech feature extraction and transmission to back-end servers.
• Limited resources on front-end devices necessitate reducing the speech-feature vector sequence frame rate.
• Previous methods adapted hidden Markov models (HMMs) individually before connecting them, leading to increased insertion errors due to state skipping.

Purpose of the Study:

• To propose and evaluate an improved model adaptation approach for distributed speech recognition.
• To address the degradation in recognition accuracy caused by frame rate mismatches in mobile speech recognition.
• To reduce insertion errors prevalent in the 'adapting-then-connecting' approach.

Main Methods:

• Introduced a 'connecting-then-adapting' approach for HMM-based speech recognition.
• This method first connects individual HMMs into a word network, then adapts the entire network's transition matrix.
• Adaptation is performed considering the target frame rate, creating optimized transition arcs.

Main Results:

• The 'connecting-then-adapting' approach significantly reduces insertion errors compared to the 'adapting-then-connecting' method.
• Experimental analysis showed improvements in insertion, deletion, and substitution error rates.
• The proposed method achieved a better overall speech recognition rate.

Conclusions:

• The 'connecting-then-adapting' approach offers superior performance for frame rate adaptation in distributed speech recognition.
• This method effectively aligns speech features with the word network states, minimizing errors.
• The findings suggest a more robust strategy for mobile speech recognition systems facing resource constraints.