Fish Audio Drama 3 in ComfyUI: directed multi-speaker voice, then H3 talking video

Fish Audio Drama 3 is a 35B-parameter text-to-speech model with 3B active per token, aimed at multi-character scenes and long-form audio. The model card lists voice cloning, natural-language control, and speaker tags from <|speaker:0|> to <|speaker:15|>, so one script can carry up to 16 characters. The weights are on Hugging Face under the Fish Audio Research License, not an OSI licence, so read it before you use the model on client work.

Benji's AI Playground's walkthrough (posted Oct 16) uses the ComfyUI-Drama3 node pack by PedroMarinhoDev (code GPL-3.0). The trick is the README's 'experts in RAM' setup. About 90% of the model is routed experts, and each audio frame uses only 8 of 256 per layer, so the experts sit in system RAM while the GPU caches the active ones. The pack lists cards from 8 GB VRAM with 32–48 GB of system RAM. It bundles Qwen3-ASR to transcribe your reference clip, and you can save an encoded voice as a small reusable file, so you don't re-run ASR for every take.

Two practical notes from the video. Inline cues like [sighs] and [laughter] mostly landed. A line that started with an ellipsis ('...') was skipped until Benji removed the dots. He then cut a roughly 55-second two-speaker scene into about 14-second chunks and fed each to MiniMax H3 with the matching character references to animate the conversation. The skysay-ai/drama3-h3-talking repo packages that idea as one ComfyUI workflow: Drama 3 voice → free VRAM → MiniMax H3 Ref2VA with the take anchored at frame 0 for lip-sync. Its demo uses a synthetic face and a public-domain reference voice.

Our take: this is the most directable open voice we've seen for drama, explainers and UGC-style dialogue, and the RAM-offload pack makes it usable on normal cards. Only clone voices you own or have explicit permission to use. We haven't run it yet. Credit Benji's AI Playground for the walkthrough, plus PedroMarinhoDev and skysay-ai for the workflows.