The role of binary mask patterns in automatic speech recognition in background noise

Area of Science:

• Speech processing
• Acoustic signal analysis
• Machine learning for audio

Background:

• Automatic Speech Recognition (ASR) performance degrades significantly in noisy environments.
• Ideal binary masks (IBMs) are known to improve ASR by separating target speech from noise.
• The specific impact of IBM patterns on ASR across varying noise conditions and vocabulary sizes remains under-explored.

Purpose of the Study:

• To investigate the role of binary mask patterns in ASR performance.
• To analyze the influence of noise, signal-to-noise ratios (SNRs), and vocabulary size on ASR with binary masking.
• To compare ASR performance with binary masking to human speech intelligibility.

Main Methods:

• Binary masks were computed using two criteria: local criterion (LC) based on SNR, and comparison of local target energy to long-term average spectral energy.
• ASR experiments were conducted under various noise conditions, SNRs, and vocabulary sizes.
• ASR performance metrics were analyzed and compared to human intelligibility data.

Main Results:

• Binary masking substantially improved ASR accuracy, even at extremely low SNRs (-60 dB).
• ASR performance patterns under different conditions qualitatively matched human speech intelligibility.
• The difference between LC and mixture SNR was a stronger predictor of ASR accuracy than LC alone.
• Optimal performance was achieved at an LC below 0 dB, contrary to maximizing SNR gain.

Conclusions:

• Binary masking is a highly effective technique for improving ASR in challenging noisy environments.
• ASR systems employing binary masks exhibit performance characteristics surprisingly similar to human listeners.
• Maximizing SNR gain is not necessarily the optimal strategy for enhancing either human or machine speech recognition in noise.