Fish Audio S2 Beta

Fish Audio S2 — Pre-Release

Best text-to-speech system among both open source and closed source.

Trained on 10M+ hours of audio across ~50 languages, S2 combines a Dual-AR architecture (Qwen3 backbone) with GRPO reinforcement learning alignment to produce natural, emotionally rich speech with fine-grained inline control.

Technical Report · Blog · Model · Playground

Model

Variant	Params	Codec	Output
S2-Pro	4B (slow) + 400M (fast)	ModifiedDAC, 10 codebooks, ~21 Hz	44.1 kHz

Highlights

• Dual-AR: Slow AR (4B) predicts semantic codebook along time axis; Fast AR (400M) fills 9 residual codebooks per step
• Inline Control: Free-form tags like [laugh], [whispers], [super happy] at word level
• RL Alignment: GRPO with unified data-reward pipeline — same model for data filtering and RL reward
• SGLang Streaming: RTF 0.195, TTFA ~100ms, 3000+ tokens/s on single H200
• 50+ Languages, multi-speaker (<|speaker:i|>), multi-turn, rapid voice cloning (10-30s reference)

What's Changed

Model & Inference

• New Dual-AR architecture with Qwen3 backbone, replacing Fish-Speech v1.5
• New ModifiedDAC audio codec (replaces Firefly/VQ-GAN)
• Support fish_qwen3_omni checkpoint format (sharded safetensors) with backward compatibility
• Fixed: torch.compile bugs, GPU memory leak, audio quality issues

Docker & Deployment

• Docker overhaul: multi-target builds, compose support, health checks, non-root user
• SGLang server integration

API & Server

• Reference voice management API (CRUD), multipart upload support
• Various server bug fixes, /health endpoint

Finetune

• Full finetune pipeline for S1/S2 (datasets, training, LoRA merge)

Docs & Infra

• README & MkDocs rewritten for S2 across 6 languages
• License updated to Fish Audio Research License
• Removed legacy code (Firefly VQ-GAN, SenseVoice, Fish Agent, old batch files)

V1.5.1

The last stable branch before the next model release.

V1.5.0

Fish Speech 1.5 release, both inference and finetune are done.

V1.4.3

Last stable release before 1.5

V1.4.2

What's Changed

• Add Audio Select to WebUI by @PoTaTo-Mika in #556
• Fix cache max_seq_len by @AnyaCoder in #568
• docs: Docker icon is missing in zh-cn README & ja README displays that it is in English & properer expression “简体中文” by @Octopus058 in #569
• docs: Corrected the wrong expressions of supported languages in README by @Octopus058 in #574
• Api json format by @AnyaCoder in #588
• Update v1.4 readmes & samples by @AnyaCoder in #592
• [chore] add docs for macos by @Tps-F in #544
• [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci in #599
• chore: typo fix on post_api by @bjwswang in #605
• feat: enable more workers in api.py by @AnyaCoder in #621
• Fix broken remove_parameterization in firefly by @med1844 in #620
• Fix dockerfile by @AnyaCoder in #622
• Fix dockerfile for pyaudio by @AnyaCoder in #623
• Update docs by @AnyaCoder in #626
• Fix backend by @AnyaCoder in #627
• Update docs by @AnyaCoder in #638

New Contributors

• @Octopus058 made their first contribution in #569
• @bjwswang made their first contribution in #605
• @med1844 made their first contribution in #620

Full Changelog: v1.4.1...v1.4.2

V1.4.1

This release includes bug fix and container optimization.

Fish Speech V1.4 Release

Fish Speech V1.4 is a leading TTS model trained on 700k hours of audio data in multiple languages.

Supported languages:

• English (en) ~300k hours
• Chinese (zh) ~300k hours
• German (de) ~20k hours
• Japanese (ja) ~20k hours
• French (fr) ~20k hours
• Spanish (es) ~20k hours
• Korean (ko) ~20k hours
• Arabic (ar) ~20k hours

Have fun :)

V1.2.1

This is the final stable release before 1.4 release on Sep 10.

Fish Speech V1.2 Release

In this release, we roll-out both 1.2 pretrain and SFT model, and also support auto-reranking for stable generation.

V1.1.2

This is the final stable release before 1.2