Noise-robust speech triage.

Area of Science:

• Speech processing
• Signal processing
• Machine learning

Background:

• Conventional speech algorithms struggle in extremely noisy environments.
• Speaker identification (SID) performance degrades significantly with low signal-to-noise ratio (SNR).
• Existing methods often require modifications to adapt to varying noise levels.

Purpose of the Study:

• To improve the performance of speech algorithms in extremely noisy environments without algorithm modification.
• To enhance speaker identification (SID), language identification, gender identification, and diarization.
• To develop robust speech activity detection (SAD) for low-SNR conditions.

Main Methods:

• Applied multiple i-vector algorithms without modification to existing speech processing pipelines.
• Pre-trained multiple speaker identification (SID) models across a range of signal-to-noise-ratio (SNR) levels.
• Implemented two optimized speech activity detection (SAD) algorithms: one for low SNR using voiced speech envelope, another for higher SNR using original speech features.

Main Results:

• Significantly improved classification accuracy (equal error rate) and processing throughput using i-vector algorithms.
• Achieved robust performance in speaker identification, language identification, gender identification, and diarization across all tested SNR levels.
• Demonstrated optimized SID performance when training and testing SNR levels were closely matched.

Conclusions:

• Unmodified conventional speech algorithms, when combined with i-vector techniques and adaptive SAD, can achieve high performance in extremely noisy conditions.
• The proposed approach offers a significant advancement in robust speech processing for real-world noisy environments.
• Optimized SAD is critical for reliable performance of downstream speech tasks when speech is barely audible.