PedroMarinhoDev / ComfyUI-Drama3

Run **Fish Audio Drama 3**, a 35B text-to-speech model for voice cloning,
dialogue and inline cues such as `[sighs]` or `[whispers]`, **natively in ComfyUI on cards from 8 GB**. 90% of the model
is routed experts and each audio frame uses only 8 of the 256 in a layer, so the experts stay in system RAM and the
GPU keeps a cache of the ones in use. The pack also brings a native **Qwen3-ASR** that writes a voice clip's
transcript for you, and nodes to save a cloned voice and load it again. Built with Fish Audio.
> [!NOTE]
> **3× faster since 2026-10-17.** The pack's own CUDA kernels read the experts the cache misses straight from RAM while
> the CPU computes others, and every layer replays as a CUDA graph: one RTX 3090 now speaks a 9 s line in 6.7 s instead
> of 16.7 s, about twice as fast as real time, two GPUs take 2.8 s, and an 8 GB card keeps pace with real time.
> Nothing new to install. Numbers.
 Drama 3 FP8 · experts in RAM · temperature 1.0, top_p 0.9, top_k 30 · RTX 3090 and RTX 4090 · no real person
cloned: the clones copy voices the model made up, the rest are voices it invents per seed. Click to listen. 
 Clone a voice from seconds of audio A clip and its transcript are all it takes; save the voice and reuse it without the codec. 
 Up to 16 speakers in one text Start each line with <|speaker:N|> and give each speaker a voice. 
 Cues where they should land [laughs] , [whispers] , [sighs] and plain-language directions inside the text ( how to write them ). 
 Long texts in one go Read in chunks the way Fish Audio's server does, each chunk hearing the ones before it. 
 Faster than real time, experts in RAM About 6 GB of the model stays on the GPU; custom CUDA kernels read the 32 GB of experts from RAM through a VRAM cache. 
 Transcripts by Qwen3-ASR Native 1.7B and 0.6B speech recognition in 30 languages, no extra packages. 
## Quick start
1. **Install the nodes.** Clone into `custom_nodes` and restart ComfyUI. There's nothing to `pip install`: the nodes
 only use what ComfyUI already has.
 [code] 
2. **Download the files** from PedroMarinhoDev/Drama-3-ComfyUI
 and PedroMarinhoDev/Qwen3-ASR-ComfyUI:
 | File | Folder | Size |
 |---|---|---|
 | `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 |
 | `qwen3_forced_aligner_0.6b_bf16.safetensors` (optional: word timing in Drama 3 Editor) | `models/asr` | 1.8 GB |
 [code] 
3. **Open the workflow and queue.** Drag `example_workflows/drama3_voice_clone.json`
 into ComfyUI, pick a few seconds of clean speech in **Load Audio**, write your text in **Drama 3 Generate**, run.
4. **Keep the voice.** Unmute **Save Drama 3 Voice** in the workflow: it writes the encoded voice (codes and transcript,
 a `.safetensors` of a few KB) to `output/drama3_voices`. **Load Drama 3 Voice** brings it back later, with no clip
 to load and nothing to encode or transcribe again.
> [!