Auden-Voice: General-Purpose Voice Encoder for Speech and Language Understanding

TL;DR

Auden-Voice is a versatile voice encoder designed to capture both speaker identity and paralinguistic cues, enhancing speech and language understanding by balancing multiple voice aspects through multi-task training.

Contribution

This work introduces Auden-Voice, a novel general-purpose voice encoder that effectively balances identity and paralinguistic features, outperforming existing methods in comprehensive evaluations.

Findings

Multi-task training yields balanced voice representations.

Contrastive language-audio pretraining improves retrieval but not paralinguistic understanding.

Auden-Voice performs well when integrated with large language models.

Abstract

Human voice encodes both identity and paralinguistic cues, yet encoders in large audio-language models (LALMs) rarely balance both aspects. In this work, we present a study toward building a general-purpose voice encoder that captures nuanced voice cues. Through a comprehensive evaluation, we find that multi-task training yields the most balanced representations, whereas contrastive language-audio pretraining (CLAP) primarily improves retrieval without enhancing paralinguistic understanding. Our final encoder, Auden-Voice, also demonstrates strong performance when integrated with LLMs. The code and training recipes will be released with the audio understanding toolkit Auden.