Expressive text to speech with voice cloning, up to 16 speakers in one text and cues like [sighs] or
[whispers]. The 32 GB of routed experts stay in system RAM and the GPU keeps a cache of the ones in use, so a
35B model runs on a consumer card. A native Qwen3-ASR writes your voice clip's transcript.
No real person is cloned here. The first four clone two voices the model made up from a single line; the rest are prompts with no voice at all, where the model invents one. Drama 3 FP8, experts in RAM, temperature 1.0, top_p 0.9, top_k 30. Copy a text to try it yourself.
Each made from one line with no reference; the cloned samples above use them like any other clip.
Three steps to the first take. You need ComfyUI v0.40 or newer, an NVIDIA GPU with 8 GB or more, and system RAM for the model file plus about 8 GB.
Clone into custom_nodes and restart ComfyUI. Nothing to pip install.
cd ComfyUI/custom_nodes git clone https://github.com/PedroMarinhoDev/ComfyUI-Drama3
From Hugging Face:
drama3_fp8experts_bf16.safetensors → models/diffusion_models 38.6 GB drama3_codec.safetensors → models/vae 1.6 GB qwen3_asr_1.7b_bf16.safetensors → models/asr 4.1 GB
Drag drama3_voice_clone.json into ComfyUI, pick a few seconds of
clean speech, write your text, run.
example_workflows/drama3_voice_clone.json
Start with FP8: Fish Audio's own expert weights, byte for byte. The format, not the GPU, sets the system RAM you need; take a smaller file when your RAM is short of the FP8 file's ~48 GB.
| File | Size | Experts | System RAM | Use it for |
|---|---|---|---|---|
drama3_fp8experts_bf16start here | 38.6 GB | FP8, as Fish Audio ships them | ~48 GB | the original weights |
drama3_w6a8experts_bf16 | 32.7 GB | W6A8 | ~40 GB | less RAM; no silent take in 16 cloned test takes |
drama3_w4a8experts_bf16 | 24.6 GB | W4A8 | ~32 GB | the least RAM; falls silent more often with real voice clips |
drama3_codec.safetensors | 1.6 GB | – | – | the original codec, needed with every file |
qwen3_asr_1.7b_bf16 / 0.6b | 4.1 / 1.6 GB | – | – | transcripts; the 0.6B takes half the time |
Milliseconds per audio frame, experts in RAM, a cloned voice, ~9 s of speech, prompt pass included; measured on 2026-10-14. Drama 3 makes 21.5 frames per second of speech: the line marks real time, 46 ms.
FP8 if you have about 48 GB of free system RAM: it's the original weights. W6A8 needs ~40 GB, W4A8 ~32 GB; W4A8 falls silent more often with real voice clips.
Only voices you have the right to use. The Fish Audio Community License forbids imitating an identifiable person without their consent.
Usually the voice: the transcript must cover every word of the clip, and the clip should be clean speech cut between sentences. Some clips make the model fall silent in Fish Audio's own code too (a known issue); try another seed or another stretch of the recording.
Keep temperature at 1.0 and top_p at 0.9: lower values add pauses and flatten the voice. The model copies the reference's delivery, so pick a lively clip for a lively take.
No. Without one the model invents a voice, a different one for each seed; the two voices on this page were made that way.
Drama 3's weights and outputs fall under the Fish Audio Community License: free for research and non-commercial use, and for commercial use below USD 500k annual revenue; a hosted TTS service needs a license from Fish Audio. Read the license for the details.