Fish Audio Drama 3: The Most Emotive Text-to-Speech Model We've Built

Drama 3 is Fish Audio's most emotive text-to-speech model: far more feeling than S2.1 Pro, and No. 1 on Artificial Analysis' Pronunciation Robustness benchmark.

October 13, 2026 | Drama 3 is available today in Text to Speech on fish.audio, the Fish Audio app (iOS and Android) and the API. Until November 12: 80% off the API and 50% off new creator plans.

The short version:

• Same direction, a stronger performance. Give Drama 3 the direction you'd give S2.1 Pro, and you get tears that sound real, anger held in check and a laugh that breaks through a straight face.
• It follows your direction. A whisper stays a whisper, a laugh lands where you put it, and the read turns where you tell it to, even halfway through a line.
• Every word still lands. Drama 3 is #1 on Artificial Analysis' independent Pronunciation Robustness benchmark: 93.2% of hard words, numbers and codes read correctly, against 88.1% for S2.1 Pro.
• Open weights. Run Drama 3 yourself from Hugging Face, free for commercial use if you and your affiliates have less than US$500,000 in annual revenue (Fish Audio Community License v1.0).

Open Drama 3 in Text to Speech →

In our launch video, Uncle Roger directs an AI clone of his own voice, made and used with his permission.

Now it's your turn. Uncle Roger's voice is on Fish Audio, so you can direct him yourself.

Direct Uncle Roger in Drama 3 →

Same Direction, Far More Feeling

Direction goes in square brackets, written the way you'd say it to an actor. Here's one line, with the same voice and the same direction, on S2.1 Pro and then on Drama 3.

S2.1 Pro

Drama 3

[barely holding it together, forcing a smile] I'm fine. [shaky breath] Really. [small laugh] [whispering] Go on without me.

Listen for the forced smile holding while the voice cracks, and for the breath, the laugh and the whisper coming exactly where they're written.

Tears, Anger, Laughter

These are the feelings speech models most often flatten. In each line, listen for the second direction: it's where the voice breaks, goes quiet or gives way to a laugh.

Choking up.

[choked up, crying quietly] I left his jacket by the door. [voice breaking] I keep thinking he'll walk in and ask for it.

Holding back anger.

[through gritted teeth, keeping his voice down] I asked you to do one thing tonight. [slow, dangerously quiet] Just one.

Laughing despite herself.

[straight-faced, trying very hard to stay serious] No, I love it, it really suits you. [cracking up despite herself] I'm sorry, it's just a very brave haircut.

It Follows Your Direction

Add new direction where the mood changes, and the read changes there.

[breathless, speaking fast] Just turn off the lights. Stay away from the windows. [voice drops to a terrified whisper] And whatever you hear out there… [slow, deadly serious] <emphasis>do not open that door.</emphasis>

The last words sit inside a paired tag, <emphasis>…</emphasis>, which directs only the words between the opening and closing tags. In Text to Speech, select the words and pick a tag from the Tags menu.

Drama 3 speaks 83 languages. Write the direction in whatever language you think in; it doesn't have to match the language of the line.

Direction in Chinese

[像是在强忍着眼泪，却努力笑着] I'm fine. Really. Go on without me.

Direction in Spanish

[conteniendo las lágrimas, pero sonriendo] I'm fine. Really. Go on without me.

Direction in Japanese

[涙をこらえながら、無理に笑って] I'm fine. Really. Go on without me.

Every Word Still Lands

More feeling only helps if the words survive it. Artificial Analysis tests that independently: its Pronunciation Robustness benchmark puts hard words in English sentences (numbers, codes, brand names, and words whose reading depends on context), and human listeners check each hard word. Drama 3 read 93.2% of them correctly, the top score on the leaderboard, against 88.1% for S2.1 Pro.

The biggest gains over S2.1 Pro are on names of brands, places and technical terms, and on words whose reading depends on context; exact sequences such as file paths and email addresses also improve, from 79.4% to 86.3%.

See the Pronunciation Robustness leaderboard →

How Drama 3 Learned to Take Direction

Speech models, our own S2 included, are usually tuned with rewards for being easy to transcribe and for sounding like the reference clip, and those scores mark a laugh or a whisper down. Drama 3 is tuned with one reward instead: a model of which take people preferred. That model also hears the script and the voice, so accuracy and likeness still count; it just doesn't punish a laugh you asked for. The technical report has the details.

What Drama 3 Is For

Drama 3 does its best work on scripted lines you render once and keep.

• Audiobooks and narration. Drama 3 can read a chapter evenly, let its voice catch on the one line that should hurt, and still get "Chapter XVII" right.
• Audio drama and fiction podcasts. A single line can go from panic to a whisper to a threat, as the door line above does.
• Games and animation. Every line of dialogue gets its own direction: anger through gritted teeth, a laugh someone can't hold in.
• Short video. Story videos and comedy skits get their tears and punchlines on cue, on the web or in the Fish Audio app.

"Voice AI has gotten very good at generating speech. The next step is directing a performance. Drama 3 is our most emotive and controllable model yet. It puts creators in the director's chair: describe what you want in plain language, listen back, and adjust only the moments that aren't working. That matters most in long-form projects like audiobooks and multilingual dubbing, where one wrong word shouldn't mean regenerating a whole chapter."

— Rissa Cao, co-founder and CEO of Fish Audio

Run Drama 3 Yourself

Like S2, Drama 3 is released as open weights, so you can run it yourself. It comes under a new license, the Fish Audio Community License v1.0: free for research and non-commercial use, and free for commercial use if you and your affiliates have less than US$500,000 in annual revenue. Above that, or to offer Drama 3 to others as a TTS, voice-cloning or inference service or API, you need a license from Fish Audio (business@fish.audio). The license also bans cloning a real person's voice without their consent. S2 stays under the Fish Audio Research License. The weights and the full license are on Hugging Face, and the code is on GitHub.

Most commercial voice models don't say how they're built. Here is how Drama 3 works. It's one large model that uses only a small part of itself at each step, with a memory suited to long scripts and several voice clips. A new audio decoder, the codebook-MTP decoder, listens back over everything so far, your voice clip included, as it fills in the fine texture of every sound. And as described above, it was tuned with a model of which takes people preferred.

For Developers and Researchers

Figure 2 from the technical report. The backbone reads the system prompt, the target text, and then its own audio frames, each entered as a fused embedding of its codes 0 and 1. The codebook-MTP decoder runs over every position on top of the backbone's hidden state. Its four shared layers take that state, the predicted code 0, and the input frame's codes 2–9, which bypass the backbone, and MTP Layers 1–9 then predict the remaining codes one at a time. Codes are predicted wherever the next token is an audio frame, every layer attends causally along time, and the audio tokenizer decoder turns the ten codes into a waveform.

• Backbone. 35B-parameter mixture-of-experts, about 3B active per token (8 of 256 routed experts plus one shared), initialized from the language model of Qwen3.5-35B-A3B.
• Hybrid attention. 40 layers in blocks of three Gated DeltaNet layers and one gated full-attention layer. The DeltaNet layers carry history in a constant-size state; only one layer in four keeps a KV cache.
• Codebook-MTP decoder. Replaces S2's per-frame Fast AR. Four shared layers feed nine full-width layers, one per acoustic codebook (1–9), and every layer attends causally over the whole sequence, including the reference audio's fine codes.
• Partial codebook fusion. Only codebooks 0 and 1 are fed back into the backbone (S2 fed back all ten), to limit exposure bias. Codes 2–9 go straight to the decoder.
• Tokenizer. S2's, unchanged and built on the Descript Audio Codec: 44.1 kHz, ten codebooks (one semantic, nine residual), about 21.5 frames per second, fully causal.
• Training. Three pre-training stages, instruction fine-tuning, then RL. Every task is written as a conversation in the same ChatML-style format.
• RL. One learned human-preference reward model and no rule-based rewards. GRPO without a KL penalty, 1,536 prompts × 8 takes per step.
• Inference. On one H200, the inference engine decodes a single stream at 271 frames per second, against 177 for S2, although Drama 3 has more than seven times as many parameters. Inference is available through SGLang-Omni.

The technical report has the full details.

Drama 3 in the API

Every script in this post works in the API as written. Set the model header to drama-3 and put the script, tags included, in text:

curl https://api.fish.audio/v1/tts \
  -H "Authorization: Bearer <your-api-key>" \
  -H "Content-Type: application/json" \
  -H "model: drama-3" \
  -d '{
    "text": "[breathless, speaking fast] Just turn off the lights. Stay away from the windows. [voice drops to a terrified whisper] And whatever you hear out there… [slow, deadly serious] <emphasis>do not open that door.</emphasis>",
    "reference_id": "536d3a5e000945adb7038665781a4aca",
    "format": "mp3"
  }' \
  -o door.mp3

It's the same endpoint as S2.1 Pro, so switching an existing integration takes one header change. Tags and parameters are in the Drama 3 docs.

Get an API key →

Coming Next

Two editing tools are in small, invite-only testing, and an editing API is coming soon. Here's a first listen.

Interactive Editing

Listen to a take, tell Drama 3 what to change, and it plays the line again with the earlier takes in mind. The direction can be relative ("softer than that one") or completely new.

The line: The results just came in. We didn't make it.

Take 1: heavy, deflated sadness

Take 2: "softer and warmer versus previous"

Take 3: "more amused than before; drop the hurt"

Audio Inpainting

You have a take you like, and then the script changes. Mark the word, type the new one, and Drama 3 regenerates just that word, fitted to the voice, mood and pace on either side. Everything else is the take you already approved. Here is the crying line from above after a one-word rewrite:

I left his jacket scarf by the door. I keep thinking he'll walk in and ask for it.

Original take

After the edit (only "scarf" is new)

Launch Offer

For developers: from October 13 to November 12, the Drama 3 API is $10 per million characters, 80% off the $50 list price.

For creators: Drama 3 costs the same credits as S2.1 Pro. Until November 12, a new Plus, Pro or Max plan is 50% off: your first month on a monthly plan, or your first year on an annual one. Comment on our launch post within the first 48 hours and we'll give you a free month of Fish Audio Plus.

Takes vary, so render a few and keep the one that plays. If a take misses your direction, post the script on Discord; that's the feedback we want most.

Bring a script. Bring your direction.

Direct your first line in Drama 3 →