arkennemasis fal · Resemble ChatterboxHD Speech to Speech · $0.02/min

Chatterbox is the model that made open TTS stop feeling like a compromise - a 0.5B Llama-backbone model, MIT licensed, zero-shot cloning from about five seconds of reference audio, and good enough that when it launched the community's question was whether it had just killed a paid subscription. This node is the other half of Chatterbox: speech to speech. You don't give it text and a voice. You give it a recording, and it re-says it in a different voice.

That distinction matters more than it looks. TTS invents a performance from your punctuation; speech-to-speech inherits one. Pauses, emphasis, the little stumble, the rise at the end of a question - all of that survives, which is exactly what you want when you have a good take and the wrong voice.

Inputs and what they do

source_audio is required - the clip to convert. Feed it a clean single-speaker recording; overlapping voices or a lot of room noise is where conversion gets mushy.

target_voice is a dropdown of the ten built-in voices: Aurora, Blade, Britney, Carl, Cliff, Richard, Rico, Siobhan and Vicky, plus (not set). Leave it unset and the model picks a random target voice, which is fun exactly once.

target_voice_audio is the interesting one: give it a sample of the voice you want and it overrides target_voice entirely. That's local-voice-cloning behaviour, except you're not running a 0.5B model and a clone stack on your own card - you're paying $0.02 per minute of audio to have fal do it. At that rate, a 30-second clip costs a penny.

high_quality_audio is off by default, which means 24 kHz output. Turn it on and you get 48 kHz and a longer wait. For anything that's going to be mixed with music or delivered as a finished track, turn it on; for scratch audio you're going to re-record anyway, don't bother.

Outputs: audio (an AUDIO, ready for the pack's assembly or a Save Audio node), audio_path, and info with the request id, the cost estimate and fal's full response.

The pack's shared controls sit underneath in the optional section: max_cost_usd (default $20), reuse_identical_run, extra_json, max_concurrent.

One thing worth knowing before you use it commercially

Chatterbox ships with built-in perceptual watermarking - Resemble builds it into the family, and the node's own description repeats it. That's a feature, not a bug, but it's the kind of feature you want to know about before a client asks. Voice cloning also inherits every obligation that comes with a person's voice; cloning someone without their permission is a legal problem regardless of which endpoint did the compute.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install httpx

Manager → search arkennemasis. Then in the .env next to run_nvidia_gpu.bat:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

Restart ComfyUI. The audio node appears under arkennemasis/fal/Audio/Chatterbox.

Gotchas

• This is not TTS. If you have words and no recording, this is the wrong node - you want a text-to-speech endpoint. Speech-to-speech needs audio in.
• Random voice if you set nothing. Both target_voice and target_voice_audio empty means the model picks, so a "why does it sound different every run" complaint is usually an unset dropdown.
• Length drives cost. The estimate is per minute of source audio, so trimming silence off the front of a long take is a real saving. reuse_identical_run (on by default) means an unchanged graph won't re-bill you, which is also why an unchanged graph won't give you a different take.
• Everything queues at max_concurrent: 1, and a failed fal node stops the queue rather than letting the rest spend.