FULL STORY

Fish Audio's Drama 3: From Open-Weights Release to Real-World Tests

Fish Audio released the open-weights Drama 3 Preview, topping open TTS leaderboards; it was quickly integrated into audio.cpp and tested by reviewers.

2026-10-14 ~ 2026-10-17 · 3 episodes · 39 posts

Episode 1 · Fish Audio Open-Sources Drama 3, Tops Open-Weight TTS Charts Past Eleven v4 (2026-10-14, 30 posts)

Fish Audio released the open-weight TTS model Drama 3 Preview on October 14, built around the idea of "directing a performance, not just reading words." The same day, Artificial Analysis's leaderboard showed it becoming the top open-weight model on the overall Text-to-Speech chart, surpassing Eleven v4 — independent data quickly validating the company's claims. Fish Audio also disclosed a $52M seed round and over 15M users.

Confirmed

  • Open-weight release: weights published in safetensors format, trending on Hugging Face
  • Natural-language directing: specify emotion, pacing, pauses and even sound effects per utterance, with precise control over which words change; the site demos 5 ways to direct the same line
  • Word-level editing & local repair: replace a single word in long audio without breaking surrounding prosody, switch voices/emotions mid-sentence, and generate multi-character dialogue in one pass
  • Architecture & performance: fishomni architecture, 35B MoE with 3B active per token, covering Chinese, English and Japanese; 271 frames/sec decoding on a single H200; multi-reference voice cloning and iterative refinement
  • Leaderboard: #1 open-weight TTS on Artificial Analysis, beating Eleven v4; two firsts and one second in Pronunciation Robustness (human-judged across four categories); demos include financial-report narration (capital-markets revenue up 7.1% QoQ to $892.2M)
  • Arena: Elo 1243 over 1,857 battles in the Provider Voice Arena, rank 10, just behind Speechify Simba 3.2, up 101 Elo over its predecessor
  • Local deployment: akhaliq's int4 quantization is 22.1GB (another post cites VRAM dropping from 37GB to 20.8GB), single-GPU runnable; @futterneid reports smooth DGX Spark performance with near-original quality at 4-bit
  • Company: 15M+ users, $52M seed round

Unconfirmed

  • One repost claims 83 languages, inconsistent with the official three-language (Chinese/English/Japanese) statement; the official figure stands

Why it matters

  • Drama 3 shifts the TTS paradigm from writing text to directing performance, making emotion and controllability the new competitive axis for open speech models
  • Open weights plus word-level editing, local repair and multi-character generation lower the barrier for audiobooks, podcasts and dubbing; @LexSokolin and others place it at the front of current TTS on expressiveness and control
  • Independent benchmarks quickly validated the claims, showing open models rapidly closing in on — or beating — leading closed commercial TTS services
  • The $52M seed round signals continued strong capital interest in speech synthesis

10 more related posts →

Episode 2 · Drama3 TTS Model Lands in audio.cpp with Up to 5x Real-Time Speed (2026-10-16, 2 posts)

audio.cpp announced support for Fish Audio's Drama3 TTS model, contributed by JarodMica, reaching up to 5x real-time speed on an RTX 5090.

Episode 3 · Fish Audio Open-Sources Drama 3: Open-Weight TTS with Natural Language Direction (2026-10-16, 7 posts)

Fish Audio released a preview of its voice model Drama 3 on October 16, with several bloggers covering and testing it around the same time. The model ships with fully open weights and supports 83 languages.

Confirmed

  • Core feature: supports natural-language "director-style" performance instructions—simply describe emotion, intent, pacing, pauses in plain text within the script, or even specify which words should change, allowing precise control over how each line is delivered rather than just reading the script verbatim
  • Supports audio inpainting
  • Covers 83 languages with fully open weights
  • Released as a Preview

Why it matters

  • mhdfaran, a blogger who says he has tested hundreds of voice models, calls Drama 3 the most emotionally expressive voice model to date, and its open weights lower the barrier to entry
  • VibeMarketer argues its value lies not just in realism but in controllability—instruction-driven tone and emotional expression turns "reading a script" into "acting," and its "speaking Japanese through it" example illustrates cross-language potential
  • The combination of natural-language emotion control and audio inpainting enables finer post-production adjustments to speech, with practical implications for dubbing and multilingual content creation