Fish Audio
13 Intel
Fish Audio is a generative AI company building text-to-speech and voice-cloning models, founded in 2025 by former Nvidia researcher Shijia Liao after he trained a voice-generation model on a single GPU and open-sourced it. Based in Mountain View, California, the company raised a $52 million seed round in 2026 led by Coreline Ventures and Capital Today, and says its open-source and hosted models are used by more than eight million people. Its research is released as open source software, including the Fish-Speech model.
Is this you? Sign in with X to claim this profile.
Intel
Fish Audio details an agent-native documentation layer: an llms.txt index, a Docs MCP server for live API lookup, and installable skills for offline code generation, aimed at preventing agents from emitting stale auth, model IDs and WebSocket schemas.
Fish Audio previewed Drama 3, a speech model directed in plain language for tone, emotion, pacing and character, with emotion shifts inside a single line. It is callable in the API as drama-3-preview. Claims come from the announcement and demo clip, not independent testing.
Fish Audio's chief scientist explains how a 4x per-GPU efficiency gain made a free TTS API viable: custom CUDA kernels, FP8 quantization, continuous batching and tighter GPU scheduling. The post gives throughput and utilization figures and open-sources its kernel library.
Fish Audio releases Drama 3, a text-to-speech model with bracketed emotion direction. It is available through the API and as open weights under a community license, and tops an independent pronunciation robustness benchmark at 93.2%.
Fish Audio launches Fish Agent, a no-code builder for voice agents that take and place calls, powered by its Live 3 full-duplex model. It adds tools, webhooks, a knowledge base, simulation tests and call analytics, priced from $0.035 per minute during a launch offer.
Fish Audio previewed Drama 3, a TTS model controllable via plain-language tone and pacing instructions, supporting mid-sentence voice shifts, multi-character scenes, and single-word corrections without audio tags.
Fish Audio's Live 3 is a full-duplex voice model that keeps listening while speaking. Trained on real conversations, it handles pauses, acknowledgments like mm-hmm, and mid-sentence changes of mind, and can decide to respond in as little as 186 ms.
Fish Audio released Drama 3 Preview, a text-to-speech model with open weights on Hugging Face and a hosted API. The vendor says it ranks first among open-weights TTS on Artificial Analysis, supports 83 languages, and costs about a fifth of Eleven v4. Claims are unverified.
Fish Audio's MCP server adds image generation and editing, video generation, lip sync and talking avatars alongside its voice tools. Agents in Claude Code, Cursor, Codex and others use one sign-in and shared credits, with a quote before each job.
Fish Audio releases S2, an open-weights text-to-speech model with natural-language inline tags for word-level control. Includes fine-tuning code and an SGLang streaming engine, reported benchmark wins on EmergentTTS-Eval, and an RL recipe reusing data-curation models as rewards.
Fish Audio launched Drama 3 in preview, a text-to-speech model controlled by bracketed natural-language directions, sound and pause tags, and paired tags that affect selected words. Multi-speaker scenes generate in one pass, and the API selects it with a model header. Quality varies between takes.
Fish Audio made its S2.1 Pro voice model free via API with unlimited Fair Use access across 83 languages, enabled by custom FP8 GEMM and FlashAttention kernels that cut latency to ~70ms time-to-first-audio and beat cuBLAS by 2-4x.
Fish Audio's Drama 3 is an open-weights 35B MoE TTS model (about 3B active) trained with a learned human-preference reward model rather than transcription and similarity rewards. It follows inline acting directions, scores 93.2% on pronunciation robustness, and works through the existing API with on