Fish Audio Drama 3: local ComfyUI setup and voice cloning

This tutorial walks through running Fish Audio Drama 3 locally in ComfyUI. The presenter describes it as a 35-billion-parameter text-to-speech model with open weights, inline emotion cues and support for up to 16 speakers in one script.

Setup requires the model weights, an encode-decode codec and Qwen3-ASR for transcription. The walkthrough places these in the diffusion models, VAE and manually created ASR folders. The presenter selects INT8 experts_BF16 weights and Qwen3-ASR 1.7B; the nodes also support the 0.6B transcription model. Experts stream from system RAM to reduce VRAM demand, but the demonstrated download is about 38 GB. The tutorial does not establish a minimum hardware specification.

The voice-cloning workflow encodes reference audio and saves reusable character voices. Loading those saved files avoids repeating transcription and encoding for each script. A two-speaker example uses numbered speaker tags and emotional cues. The presenter reports that leading ellipses caused several lines to disappear, then demonstrates a complete take after removing them. This is an observed fix in the demonstrated workflow, rather than a universal punctuation rule.

The final section passes the generated dialogue to MiniMax H3 in roughly 13- to 14-second clips to animate the two characters.