Fish Audio S2 combines unified data pipelines with dual-autoregressive generation for controllable, expressive text-to-speech.
We introduce Fish Audio S2, an open-sourced text-to-speech system featuring multi-speaker, multi-turn generation, and, most importantly, instruction-following control via natural-language descriptions. To scale training, we develop a multi-stage training recipe together with a staged data pipeline covering video captioning and speech captioning, voice-quality assessment, and reward modeling. To push the frontier of open-source TTS, we release our model weights, fine-tuning code, and an SGLang-based inference engine. The inference engine is production-ready for streaming, achieving an RTF of 0.195 and a time-to-first-audio below 100 ms.Our code and weights are available on GitHub (https://github.com/fishaudio/fish-speech) and Hugging Face (https://huggingface.co/fishaudio/s2-pro). We highly encourage readers to visit https://fish.audio to try custom voices.
Fish Audio S2 uses a unified data pipeline where quality assessment and ASR models serve dual purposes: filtering training data and providing RL reward signals, eliminating distribution mismatch.
The Dual-Autoregressive architecture separates linguistic planning via a Slow AR from acoustic detail generation via a Fast AR, solving the dimensionality problem of ten stacked codebooks.
Fish Audio S2 achieves superior instruction-following, multi-speaker multi-turn generation, and stable long-form synthesis with production-ready streaming at 0.195 RTF and sub-100ms latency.
The paper mentions that generating fine-grained natural-language instructions for vocal features at scale remains a major bottleneck in TTS development.
The script cuts off mid-sentence during the Fast AR architecture explanation, preventing complete assessment of all technical limitations discussed.
Yuejie Li, Ke Yang +5
Spoken query retrieval is an important interaction mode in modern information retrieval. However, existing evaluation datasets are often limited to simple queries under constrained noise conditions, m...
Yitian Gong, Kuangwei Chen +10
Discrete audio tokenizers are fundamental to empowering large language models with native audio processing and generation capabilities. Despite recent progress, existing approaches often rely on pretr...
Hangrui Hu, Xinfa Zhu +14
In this report, we present the Qwen3-TTS series, a family of advanced multilingual, controllable, robust, and streaming text-to-speech models. Qwen3-TTS supports state-of-the-art 3-second voice clonin...
Georgii Aparin, Tasnima Sadekova +6
Sparse Autoencoders (SAEs) are powerful tools for interpreting neural representations, yet their use in audio remains underexplored. We train SAEs across all encoder layers of Whisper and HuBERT, prov...
Xian Shi, Xiong Wang +11
In this report, we introduce Qwen3-ASR family, which includes two powerful all-in-one speech recognition models and a novel non-autoregressive speech forced alignment model. Qwen3-ASR-1.7B and Qwen3-A...
