Continuous-Token Diffusion for Speaker-Referenced TTS in Multimodal LLMs

2025-10-14

AI Summary

Existing MLLM-based TTS approaches rely on discrete speech tokens, neglecting speech continuity and consequently suffering fine-grained acoustic information loss. To address this, we propose the first end-to-end MLLM-TTS framework explicitly designed for continuous speech representation. Our method introduces a dual-head architecture—comprising a language modeling head and a frame-level diffusion head—along with a novel continuous-token diffusion mechanism. We adopt a two-stage training strategy: first freezing the language model head to optimize the diffusion head, then jointly fine-tuning both heads; masked training is further incorporated to mitigate exposure bias. Evaluated on LibriSpeech test-clean, our approach achieves a word error rate (WER) of 1.95%—a 46% relative reduction over single-stage training—along with a speaker similarity score of 0.54 and an UTMOS of 4.00. To the best of our knowledge, this is the first work to realize high-fidelity, controllable continuous speech synthesis within the MLLM paradigm.

Technology Category

Machine Learning: Large Multimodal Models (LMMs)Natural Language Processing: (Large) Language ModelsPlanning, Routing, and Scheduling: Planning with Language Models

Application Category

User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendationSearch and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved informationSemantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Unified architectures in multimodal large language models (MLLM) have shown promise in handling diverse tasks within a single framework. In the text-to-speech (TTS) task, current MLLM-based approaches rely on discrete token representations, which disregard the inherently continuous nature of speech and can lead to loss of fine-grained acoustic information.In this work, we investigate the TTS within the MLLM paradigm using continuous speech representations. We design a dual-head architecture and implement two complementary training strategies for a robust model. (1) A diffusion head generating continuous speech representations is added on the MLLM, which is on frame-level and strictly autoregressive. (2) The original language model head is retained to preserve multitask capability and to control the start and end of speech synthesis. (3) Masked training is employed to address exposure bias in autoregressive decoding. (4) To stabilize optimization, we propose a two-stage scheme where the LM is frozen in the second stage, ensuring the diffusion head learns from a fixed input distribution. Evaluations on LibriSpeech(PC) test-clean show that our approach achieves state-of-the-art autoregressive performance, with a WER of 1.95%, speaker similarity of 0.54, and UTMOS of 4.00. The two-stage training yields a 46% relative WER reduction over the one-stage training baseline. These results highlight the effectiveness of combining autoregressive modeling with continuous-token diffusion, supported by a two-stage training procedure.

Problem

• Addressing loss of acoustic detail in text-to-speech systems using continuous representations
• Improving speaker similarity and speech quality in multimodal language models
• Stabilizing training for autoregressive speech synthesis with diffusion models

Innovation

• Continuous speech representations replace discrete tokens
• Dual-head architecture with diffusion and language model heads
• Two-stage training stabilizes optimization and improves performance