Utterance Clustering Using Stereo Audio Channels.

Area of Science:

• Audio signal processing
• Machine learning
• Speech recognition

Background:

• Utterance clustering is crucial for separating speakers in audio.
• Current methods often rely on single-channel (mono) audio signals.
• Improving clustering performance in complex acoustic environments remains a challenge.

Purpose of the Study:

• To enhance utterance clustering performance by utilizing multichannel (stereo) audio signals.
• To investigate novel methods for processing stereo audio for improved feature extraction.
• To evaluate the effectiveness of the proposed approach against conventional mono-signal methods.

Main Methods:

• Processed stereo audio signals by combining left and right channels.
• Extracted embedded features, known as d-vectors, from processed audio.
• Applied a parameter-sharing Gaussian mixture model for supervised utterance clustering.
• Utilized maximum likelihood for speaker identification during testing.

Main Results:

• The proposed multichannel audio processing method significantly improved utterance clustering performance.
• Experimental results demonstrated superior accuracy compared to conventional mono-audio signal methods.
• The method showed effectiveness even in complex multiperson discussion scenarios.

Conclusions:

• Multichannel audio signal processing offers a significant advantage for utterance clustering.
• The developed d-vector extraction and Gaussian mixture model approach is effective.
• This research provides a more robust solution for speaker diarization in challenging audio conditions.