Underwater short-utterance speaker identification via Active Identity Embedding and parameter-efficient adaptation.

Area of Science:

• Underwater acoustics
• Speech processing
• Biometric security

Background:

• Diver safety in emergencies relies on effective speaker identification.
• Traditional methods struggle with short distress calls and underwater distortion.
• Acoustic-phonetic features are degraded by severe underwater channel distortions.

Purpose of the Study:

• To develop a robust speaker identification system for underwater emergency scenarios.
• To overcome limitations of passive voiceprint extraction for short and distorted speech.
• To propose a novel Active Identity Embedding approach.

Main Methods:

• Reformulated speaker identification as an active message embedding and extraction task.
• Integrated identity embedding/extraction into HiFi-Codec architecture with a multi-dilation Gated Convolutional Unit decoder.
• Employed a two-stage pretraining-fine-tuning strategy with a hybrid scheme (low-rank adaptation, full-parameter fine-tuning, parameter freezing).

Main Results:

• Successfully embedded a 16-bit identity message into 1-second speech segments with minimal quality degradation (PESQ-NB ≤ 0.07).
• Achieved a bit error rate < 1% at 15 dB signal-to-noise ratio.
• Reduced trainable parameters to 18.7% of full-parameter fine-tuning via the hybrid scheme.

Conclusions:

• Active Identity Embedding offers a paradigm shift for robust underwater speaker identification.
• The proposed HiFi-Codec integration and fine-tuning strategy effectively handle distorted and short speech.
• This approach enhances diver safety by enabling reliable communication in challenging environments.