The impact of exploiting spectro-temporal context in computational speech segregation

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Computational speech segregation aims to separate speech from noise using ideal binary mask estimation.
• Existing methods exploit spectro-temporal context via delta features or support vector machine (SVM) integration to enhance mask accuracy.

Purpose of the Study:

• To investigate the effectiveness of delta features and SVM integration for speech segregation.
• To analyze the impact of spectro-temporal context on mask estimation accuracy.
• To evaluate system generalization to novel noise segments and compare objective intelligibility measures.

Main Methods:

• Two experiments were conducted to assess speech segregation strategies.
• Experiment I: Evaluated delta features and SVM integration in stationary and six-talker noise.
• Experiment II: Focused on delta features with novel noise segments, analyzing intelligibility and objective measures.

Main Results:

• Delta features yielded higher speech intelligibility compared to SVM integration.
• Intelligibility increased with the amount of spectral information utilized by delta features.
• The system demonstrated poor generalization to novel noise segments.
• Objective measures (e.g., extended short-term objective intelligibility) did not fully correlate with measured intelligibility.

Conclusions:

• Delta features are a promising strategy for improving speech segregation, outperforming SVM integration.
• Current systems require further development for robust generalization to unseen noise conditions.
• The findings suggest a need for improved cost functions that better correlate with perceived speech intelligibility in noise.