DiffuSpeech: Silent Thought, Spoken Answer via Unified Speech-Text Diffusion
This addresses the issue of uncorrectable errors in speech generation for users of speech AI systems, presenting a novel paradigm rather than an incremental improvement.
The paper tackles the problem of speech language models generating errors without explicit reasoning by introducing a 'Silent Thought, Spoken Answer' paradigm, where internal text reasoning is generated alongside spoken responses, resulting in state-of-the-art speech-to-speech QA accuracy with up to 9 points improvement over baselines and high TTS quality (6.2% WER).
Current speech language models generate responses directly without explicit reasoning, leading to errors that cannot be corrected once audio is produced. We introduce \textbf{``Silent Thought, Spoken Answer''} -- a paradigm where speech LLMs generate internal text reasoning alongside spoken responses, with thinking traces informing speech quality. To realize this, we present \method{}, the first diffusion-based speech-text language model supporting both understanding and generation, unifying discrete text and tokenized speech under a single masked diffusion framework. Unlike autoregressive approaches, \method{} jointly generates reasoning traces and speech tokens through iterative denoising, with modality-specific masking schedules. We also construct \dataset{}, the first speech QA dataset with paired text reasoning traces, containing 26K samples totaling 319 hours. Experiments show \method{} achieves state-of-the-art speech-to-speech QA accuracy, outperforming the best baseline by up to 9 points, while attaining the best TTS quality among generative models (6.2\% WER) and preserving language understanding (66.2\% MMLU). Ablations confirm that both the diffusion architecture and thinking traces contribute to these gains.