Model adaptation method for recognition of speech with missing frames

Area of Science:

• Speech Recognition
• Signal Processing
• Information Theory

Background:

• Data packet loss is a significant challenge in distributed speech recognition (DSR) over error-prone networks.
• Linear interpolation is a common but potentially suboptimal method for reconstructing lost data frames.
• Robustness and computational efficiency are critical for real-world DSR systems.

Purpose of the Study:

• To propose a novel error-concealment decoding method for DSR that dynamically adapts hidden Markov model (HMM) transition probabilities.
• To evaluate the performance of the proposed method against traditional data reconstruction techniques.
• To assess the robustness and computational efficiency of the new approach under varying frame loss conditions.

Main Methods:

• Developed an error-concealment decoding technique that adjusts HMM transition probabilities based on observed frame loss.
• Implemented and tested the proposed method within a DSR system.
• Compared the proposed method with linear interpolation for data reconstruction.

Main Results:

• The proposed error-concealment method achieved comparable accuracy to linear interpolation.
• The system demonstrated enhanced robustness against heavy frame loss.
• Significant reductions in computation time were observed compared to the data reconstruction method.

Conclusions:

• The dynamic adaptation of HMM transition probabilities offers an effective error-concealment strategy for DSR.
• This approach provides a more robust and computationally efficient alternative to linear interpolation for handling frame loss.
• The findings suggest practical benefits for deploying DSR in challenging network environments.