Make a two-speaker dialogue WAV with Fish Audio Drama 3

Fish Audio Drama 3 is a multi-speaker TTS model that turns a script into speech. Add speaker tags and dialogue, then provide a reference audio clip and its transcript for each voice to create one WAV with multiple speakers. It fits scenes or longer readings with recurring voice changes. Drama 3 model card

Drama 3 is a mixture-of-experts model with 35B total parameters and 3B active for each token⁴. Its model card lists 83 languages, and the script can identify up to 16 speakers with tags from speaker 0 through speaker 15. Repeat a speaker's tag on each turn so the same character keeps the assigned voice.

Speaker tags identify who reads each line, while short bracketed cues describe a moment's delivery. The model card shows point tags such as [sighs] and [laughs]. Generate the passage and listen to whether the cue produces the delivery you intended.

Fish Audio's current installation guide requires at least 48 GB of GPU memory for Drama 3 FP8. It supports NVIDIA GPUs on Linux or WSL. CPU⁵, Apple MPS, AMD ROCm⁶, and torch.compile are not supported, so choose another model if you need to generate on a Mac or CPU. Official installation requirements

The install guide lists 24 GB of GPU memory for S2-Pro and supports CPU or NVIDIA GPU inference⁷. Drama 3's FP8 path needs at least 48 GB of NVIDIA GPU memory, so the two models target different hardware.

The published run used one concurrent stream on an H200 and a pre-encoded English Seed-TTS reference; the result aggregates four runs. The mean RTF⁸ and TTFA p50 are results for this setup.

On Ubuntu or WSL, prepare Python 3.12 and an NVIDIA driver, then install the audio system packages listed by the repository. The commands below clone the source and install its CUDA⁹ 12.9 dependencies into the Python environment.

sudo apt-get update && sudo apt-get install -y portaudio19-dev libsox-dev ffmpeg
curl -LsSf https://astral.sh/uv/install.sh | sh
source "$HOME/.local/bin/env"
git clone https://github.com/fishaudio/fish-speech.git
cd fish-speech
uv sync --python 3.12 --extra cu129

After uv sync finishes, keep working from the repository directory. If your shell is elsewhere, change into fish-speech before continuing.

Download this Drama 3 revision into checkpoints/drama-3. It includes the codec¹⁰.pth required for inference, so you do not need separate S2-Pro weights. The public model can be downloaded without a login token.

uv pip install "huggingface_hub[cli]"
uv run hf download fishaudio/drama-3 --revision e0a9218aa48888270f06ef1ed9ea67622c0e17d2 --local-dir checkpoints/drama-3

Save the two reference recordings as speaker0.wav and speaker1.wav in the fish-speech directory. Listen to each clip and transcribe its exact words. Use recordings with the speakers' permission.

The CLI¹¹ pairs reference audio and transcripts in the order you provide them. Put the exact words spoken in speaker0.wav after that clip, then the transcript for speaker1.wav. The lines below are examples; replace them with the actual words in your WAV files before running the command. Keep the script's speaker numbers in the same order.

The command pairs two reference WAVs with their transcripts, then switches speakers within one script. The official CLI accepts repeated reference-audio and transcript options. Official Drama 3 CLI

uv run python fish_speech/models/text2semantic/inference.py \
  --checkpoint-path checkpoints/drama-3 \
  --text "<|speaker:0|>제가 먼저 장면을 설명할게요. [laughs] <|speaker:1|>좋아요. 다음 대사를 이어가죠. [sighs]" \
  --prompt-audio speaker0.wav \
  --prompt-text "오늘은 새 전시를 소개합니다." \
  --prompt-audio speaker1.wav \
  --prompt-text "첫 번째 작품은 창가에 걸려 있어요." \
  --output drama-two-voices.wav \
  --device cuda:0

When the command completes, drama-two-voices.wav appears in the fish-speech directory. Listen through it to check voice changes, each line's pronunciation, and the placement of the laugh and sigh. If a voice sounds wrong, check that speaker's reference clip and transcript before generating again.

The example selects cuda:0 for one GPU. For multiple GPUs, pass the CLI's comma-separated device list and reserve memory on the last GPU for both the tower and codec.

Drama 3 ships with the Fish Audio Community License 1.0, updated October 13, 2026. It permits research and non-commercial use, plus limited commercial use when your and your affiliates' combined gross revenue from all sources over the preceding 12 months is below USD 500,000. If you install the model for a third party's operations or deliver them a copy of the model, that company must also meet the threshold. Full license

A self-service API¹², website, or app where users request speech generation is a Hosted Service and needs a separate written license regardless of revenue or purpose. The license excludes speech for content or products you create or control, and production work where you operate the model and deliver only finished files. A free, non-commercial research or demo service is also excluded when it meets the stated conditions. Distribution requires the “Built with Fish Audio” attribution and a license copy; obtain consent before imitating an identifiable person's voice.

Checked October 16, 2026: Hugging Face metadata still shows the earlier Research License label. The conditions above follow the LICENSE text distributed with the current model.