Fish Audio Drama 3 Is Free to Try: The Open-Weights AI Voice Model You Direct Like an Actor (And the 48 GB Catch)

Fish Audio just released Fish Audio Drama 3, a text-to-speech model built for scenes rather than single lines. You can clone a voice from a few seconds of audio, cast up to 16 characters in one script, and give the model notes like “gentler, and a bit sadder” to get a new take instead of starting over. The weights went up on Hugging Face on October 13, and there’s a free demo you can use in your browser today.

For creators, it’s the first open-weights voice model that feels like directing an actor. The catch: running it at home needs a 48 GB NVIDIA card. Here’s how to use Fish Audio Drama 3 free right now, what the license allows, and where independent testing puts it next to ElevenLabs.

Fish Audio Drama 3 at a Glance

• What it is: A 35B-parameter mixture-of-experts TTS model (3B active per token) with voice cloning, inline emotion tags, multi-character scenes and 83 languages.
• What’s free: The official Drama 3 Hugging Face Space, plus the open weights (41.5 GB) if you have the hardware.
• What it costs: The hosted API lists at $50 per million characters, with a launch price of $10 per million until November 12, per Fish Audio’s announcement.
• Independent ranking: #10 overall and the top open-weights model on the Artificial Analysis TTS arena (Elo 1243). That’s still about 85 points behind Eleven v4.
• The catch: Running it locally officially needs 48 GB+ of NVIDIA VRAM, and you can only use it commercially while your business earns under $500K a year.

What Fish Audio Drama 3 Actually Does Differently

Most TTS models give you one take per request. If the read is wrong, you reword the prompt and hope. Drama 3 adds several things the Fish Speech repo calls new in this generation:

• Multi-turn director revision. Send the previous take back with a note and the model revises it in context. For voiceover work, that’s closer to how a session with a real actor goes.
• Editing as continuation. Change “Wednesday” to “Thursday” in a finished line and the model regenerates just that part, with timbre and transitions that match. The model card shows this with a previous_takes field in the API request. Fish says the interactive editing UI and word-level inpainting are still invite-only.
• Scene casting. Speaker tags from <|speaker:0|> to <|speaker:15|> let one script carry up to 16 characters. Long text is chunked and can be streamed.
• Plain-language direction. Bracket tags like [whispering], [sighs] or [choked up, crying quietly] go right where you want them in the line. You can write the direction in a different language from the dialogue.

Architecture details are in the technical report.

Is It Really Better? What Was Measured and by Whom

Here’s what was independently measured and what wasn’t.

• Independent (Artificial Analysis): On the blind-vote TTS arena, “Fish Audio Drama 3 Preview” scores 1243 ±16, rank range 8–10. That puts it above open-weights rival Breeze TTS 2 (1221, rank 11–16) and well above Fish’s own S2 Pro (1118). It does not beat ElevenLabs: Eleven v4 and v4 Turbo sit at 1328–1330, a gap that is clearly larger than the margin of error. Fish also cites a first-place 93.2% on Artificial Analysis’s Pronunciation Robustness test (numbers, codes, brand names).
• Vendor-reported: Speed numbers come from Fish’s own H200 tests. The paper claims 271 frames per second for a single stream, while the model card’s serving table shows 166.7 frames/s with sub-100 ms time-to-first-audio under a different setup. Claims that Drama 3 is the “most emotive” model rely on demos, not a published benchmark.

Fish Audio Drama 3 vs. the Alternatives

We covered Breeze TTS 2 when it was the top open-weights voice. Drama 3 now edges past it, and its license is friendlier to paid work. Here’s how they compare:

Need the leaderboard’s top spot? Our Eleven v4 breakdown covers its free tier.

Try Fish Audio Drama 3 Free Today

Option 1: The browser demo (no install, no GPU). Open the official fishaudio/Drama-3 Space. Upload or record a 0.5–30 second clip of one speaker, paste exactly what that clip says into the transcript box, type your new line, and click Generate. Then download the result from the audio player. It runs on ZeroGPU with a daily allowance for free Hugging Face accounts. It’s single-speaker only; multi-character scenes need the API or a local install.

Option 2: Fish Audio’s web app. The free plan includes 8,000 credits a month (about 7 minutes of audio) at up to 500 characters per generation. Its FAQ limits free-plan output to personal, non-commercial projects. The pricing page doesn’t say whether free accounts get Drama 3, so the Space is the sure path.

Option 3: Run it locally (if you have the GPU). According to the official install guide, you need Linux or WSL and an NVIDIA GPU with at least 48 GB of VRAM (FP8). CPU, Apple MPS and AMD ROCm are not supported. Install and download:

git clone https://github.com/fishaudio/fish-speech.git && cd fish-speech
apt install portaudio19-dev libsox-dev ffmpeg
uv sync --python 3.12 --extra cu129
pip install -U "huggingface_hub[cli]"
hf download fishaudio/drama-3 --local-dir checkpoints/drama-3

Then generate a cloned line in one command (see the inference docs):

python fish_speech/models/text2semantic/inference.py \
  --checkpoint-path checkpoints/drama-3 \
  --text "[sighs] Fine. If we miss it, it's on you." \
  --prompt-audio reference.wav --prompt-text "What the reference clip says" \
  --output output.wav --device cuda:0

Prefer a GUI? The fishaudio/fish-speech:server-cuda Docker image includes a WebUI at localhost:8080/ui. For streaming and multi-speaker scenes, Fish recommends the SGLang-Omni Drama 3 cookbook (sgl-omni serve --model-path fishaudio/drama-3 --port 8000). Community 4-bit builds, such as a 25.9 GB MLX Q4 for Apple Silicon, already exist but are unofficial and untested by us.

Honest Trade-Offs Before You Spend an Afternoon

• Hardware is the real barrier. 48 GB of VRAM means an RTX 6000-class card or a rented cloud GPU. A 24 GB RTX 4090 won’t run the official build.
• It’s a pre-release. The GitHub release is tagged v3.0.0-beta. The docs say --compile isn’t supported, the float16 path for older GPUs hasn’t been tested on real hardware, and no fine-tuning code ships yet.
• The license has limits. Under the Fish Audio Community License, commercial use is free only while you and your affiliates earn under $500K a year. Running it as a hosted TTS or voice-cloning service for others always needs a separate deal. Delivering finished voiceovers to clients is explicitly allowed. Cloning a real person’s voice without consent is banned.
• Still behind the paid leaders. Arena voters prefer ElevenLabs, Qwen-Audio 3.1 and Sonic 3.6, and Fish itself suggests rendering several takes.

Workflow Tips for Drama 3

• Make the transcript match the clip word for word. The Space uses it to align the voice, and a sloppy transcript makes cloning less accurate.
• Put tags exactly where the moment happens. [laughs] lands where you type it, so write tags into the line like stage directions, not at the start of the sentence.
• Revise, don’t re-roll. If you’re on the API, send the bad take back as previous_takes with a note. It keeps the character consistent across fixes.
• Prototype free, render later. Lock your voice and tag style in the Space before renting a 48 GB cloud GPU.
• Keep a CPU fallback. For scratch narration and timing passes, Kokoro-82M or the browser-based tool from our Sopro V2 Turbo guide is fast enough and costs nothing.

We’ve followed Fish Audio since its S1 model. Our Fish.Audio S1 launch video shows how far emotional TTS has come since then, and more voice and video tests are on the Creative AI Show YouTube channel.

Fish Audio Drama 3 FAQ

Is Fish Audio Drama 3 free?

Yes, in two ways. The official Hugging Face Space is free to use with a Hugging Face account (within ZeroGPU limits), and the weights are free to download. The hosted Fish Audio API is paid, at $10 per million characters during the launch promotion and $50 list price after that.

Can I use Fish Audio Drama 3 commercially?

Yes, if your annual revenue (including affiliates) is under $500,000, and you aren’t offering it to others as a TTS or voice-cloning service. Freelancers and studios can deliver generated audio to clients of any size. Above the threshold, you’ll need a license from Fish Audio.

What are Fish Audio Drama 3’s system requirements?

The official build needs Linux or WSL, Python 3.12, and an NVIDIA GPU with at least 48 GB of VRAM. The download is 41.5 GB. Macs, AMD GPUs and CPU-only machines aren’t officially supported.

Is Fish Audio Drama 3 better than ElevenLabs?

Not on overall preference: Artificial Analysis’s arena puts Eleven v4 about 85 Elo points ahead. Drama 3 is, however, the highest-ranked model you can download and run yourself.

Accurate as of October 14, 2026. Pricing, promotions and leaderboard positions move fast.