TokenSplit: Using Discrete Speech Representations for Direct, Refined, and Transcript-Conditioned Speech Separation and Recognition

TL;DR

TokenSplit is a versatile Transformer-based model that performs speech separation, transcription, and speech synthesis directly on discrete token sequences, demonstrating high-quality results through multi-task training and refinement techniques.

Contribution

The paper introduces TokenSplit, a novel sequence-to-sequence model operating on discrete tokens for simultaneous speech separation, transcription, and synthesis, with a refinement variant improving audio quality.

Findings

Achieves state-of-the-art separation performance with or without transcript conditioning.

Demonstrates high-quality speech recognition accuracy.

Provides effective speech synthesis from discrete tokens.

Abstract

We present TokenSplit, a speech separation model that acts on discrete token sequences. The model is trained on multiple tasks simultaneously: separate and transcribe each speech source, and generate speech from text. The model operates on transcripts and audio token sequences and achieves multiple tasks through masking of inputs. The model is a sequence-to-sequence encoder-decoder model that uses the Transformer architecture. We also present a "refinement" version of the model that predicts enhanced audio tokens from the audio tokens of speech separated by a conventional separation model. Using both objective metrics and subjective MUSHRA listening tests, we show that our model achieves excellent performance in terms of separation, both with or without transcript conditioning. We also measure the automatic speech recognition (ASR) performance and provide audio samples of speech synthesis to demonstrate the additional utility of our model.