MMM: Multi-Layer Multi-Residual Multi-Stream Discrete Speech Representation from Self-supervised Learning Model

TL;DR

This paper introduces MMM, a novel multi-layer, multi-residual, multi-stream discrete speech representation method from self-supervised learning models, improving performance in speech tasks over existing discrete units.

Contribution

The paper presents MMM, a new approach combining iterative residual vector quantization and K-means to extract multi-stream discrete speech units from SSL models, outperforming prior discrete representations.

Findings

MMM surpasses neural codec performance in speech tasks.

Multi-stream discrete units improve speech recognition and synthesis.

The method achieves comparable or better results than spectral features.

Abstract

Speech discrete representation has proven effective in various downstream applications due to its superior compression rate of the waveform, fast convergence during training, and compatibility with other modalities. Discrete units extracted from self-supervised learning (SSL) models have emerged as a prominent approach for obtaining speech discrete representation. However, while discrete units have shown effectiveness compared to spectral features, they still lag behind continuous SSL representations. In this work, we propose MMM, a multi-layer multi-residual multi-stream discrete units extraction method from SSL. Specifically, we introduce iterative residual vector quantization with K-means for different layers in an SSL model to extract multi-stream speech discrete representation. Through extensive experiments in speech recognition, speech resynthesis, and text-to-speech, we demonstrate the proposed MMM can surpass or on-par with neural codec's performance under various conditions.