Drama 3

Drama 3

Speaks a line you write in a voice you choose or design, and lets you direct each take with notes, using the Fish Audio Drama 3 model. For creators exploring expressive speech and voice cloning.

Made by fishaudio on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Drama 3 is a Gradio page from Fish Audio for directing spoken performances. You type a line, choose or design a voice, and shape how it is delivered. Each take runs the Drama 3 model on ZeroGPU, with no inference endpoint or endpoint token.

Who it suits: It suits creators who want expressive speech they can steer, and developers who want to see how a large speech model is hosted on ZeroGPU.

Industry | Who uses it | How they use it | Example
Media | Video and audio creators | Draft spoken lines with a chosen voice | Type a line and listen to the take
Gaming | Game writers | Try how a character might sound | Describe a voice and cast it
Entertainment | Writers and directors | Hear a script line and adjust the delivery | Ask for another take with a new note
Software | Python developers | Study a streaming ZeroGPU speech app | Read app.py, zerogpu_backend.py and stream_transport.py

Skip it if you need long, polished productions, since the page works one take at a time and keeps no history after a restart.

Where to use it:

• A web browser, on the Space's page at Hugging Face, which runs on ZeroGPU
• Your own copy of the Space on ZeroGPU hardware
• Your own Windows, macOS or Linux computer with Python 3.12

Use case | Who | What they get back
Voice a line | Creators | A take streamed back as the GPU produces it
Cast a voice from words | Game writers | A new voice made from your description
Redirect a performance | Directors | Another take in the same voice with new direction

How to set it up: The online page needs nothing. A local copy needs Python 3.12 and the packages in requirements.txt. The real engine also needs a GPU with CUDA dependencies.

• Python 3.12 and pip
• Python packages from requirements.txt
• System packages from packages.txt, which the Space installs with apt-get
• Windows, macOS or Linux

1. Install the Hugging Face hub library.
2. Download the Space's files into a folder named Drama-3.
3. Move into that folder.
4. Install Gradio 6.28.0, then the packages from requirements.txt.
5. Start app.py.

pip install -U huggingface_hub
hf download fishaudio/Drama-3 --repo-type space --local-dir Drama-3
cd Drama-3
pip install gradio==6.28.0
pip install -r requirements.txt
python app.py

To run your own copy on Hugging Face, duplicate the Space with ZeroGPU hardware. No secrets or OAuth setup are needed. The build preloads a checkpoint of about 41 GB, so the first startup takes several minutes.

How to use it: Pick a mode, give it a line, and play the take. Each mode has its own job.

Tool | What it does
Surprise | The first mode, selected by default
Clone | Starts from one of three preset voices, a clip you upload or your microphone, and the transcript stays editable
Design | Makes a voice from a written description
Give a note | Makes another take in the same voice with new acting direction
Workflow | A canvas tab that uses the same generation code

These optional variables tune a copy that you host.

Variable | What it does
ZEROGPU_SIZE | Picks the GPU allocation. The default xlarge is the full GPU, and large asks for less and uses less quota each second
DRAMA3_WARM_UP | Defaults to 1, and 0 skips the GPU warmup at startup
ZEROGPU_COLD_START_S | Defaults to 20, the GPU time allowed for a fresh worker
DRAMA3_COMPILE | Defaults to 1, and 0 turns off regional compilation
WHISPER_MODEL | Defaults to turbo, the model that transcribes reference clips on CPU

Example: You type a line, choose one of the three preset voices in Clone mode and play the take. You then give a note such as a new mood and listen to a second take in the same voice.

For newcomers and experts: Start with Surprise, the default mode, and play one take before trying the others. The first audio may take longer while a GPU is allocated.

Each take runs the real Drama 3 model inside a GPU call, and audio streams back as the GPU makes it. Cancelling closes the generator at the next audio window. The CPU front end passes each visitor's request context to the GPU call, so Hugging Face applies that visitor's quota. The checkpoint keeps FP8 experts and bfloat16 weights, with temperature 1 and top-p 0.9 by default. For tests without a GPU, the repository offers a preview requirements file and a DEMO_PREVIEW switch.

You type a line and pick a mode
The CPU front end passes your quota context
Drama 3 runs in a ZeroGPU call
Audio streams back as it is made
You play it or ask for another take

Good to know:

• Limits: A caller may run two takes together and start 60 in any ten minutes. Those guard against bursts and do not cap a day.
• Quota: Each visitor spends their own ZeroGPU quota and gets Hugging Face's queue priority. Quota errors show up on the performance stage.
• Memory: Your reference clips, takes and direction history live only in memory, so a restart wipes them.
• License: apache-2. The fish_speech folder keeps its own license.
• Cost to run: none named for visitors beyond their ZeroGPU quota. A duplicate needs ZeroGPU hardware and no secrets.
• Data it reaches: Your text, descriptions, uploaded clips and microphone audio. Whisper transcribes reference clips on CPU.
• Not for: Cloning voices you have no permission to use.

On Hugging Face, 13 November 2026

• Likes: 166
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/fishaudio/Drama-3
• Files: https://huggingface.co/spaces/fishaudio/Drama-3/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.