kinescript-chatterbox

Chatterbox voices for kine voice, from Resemble AI's open models. It speaks the text of a scene's sounds into WAV files on your own machine and writes the time of each word, so layers can start on a word and captions follow the voice. Chatterbox Turbo acts on tags such as [laugh] and [sigh], and either model can copy the voice of a short clip.

{"id": "vo1", "role": "voice", "backend": "chatterbox",
 "text": "Well [sigh] that render took a while. [[go]][laugh] But look at it."}

kine voice scene.json     # writes voice/vo1.wav and voice/vo1.words.json
kine check scene.json

Install

It needs Kinescript 1.3 or newer, and Chatterbox needs PyTorch. Into the same environment as Kinescript:

pip install kinescript-chatterbox

From a checkout of Kinescript, pip install ./plugins/kinescript-chatterbox does the same.

On Python 3.10 to 3.13, Chatterbox pins PyTorch 2.6, whose CUDA builds stop before the RTX 50 series, so on such a card it runs on the CPU. On Python 3.14 it takes PyTorch 2.9 or newer; install a CUDA build first to use the GPU:

pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install kinescript-chatterbox

To keep it out of Kinescript's environment, install it into another Python and point KINE_VOICE_CMD_CHATTERBOX at it; kine voice sends it the lines as JSON on stdin and never through a shell:

python -m venv ~/chatterbox-env
~/chatterbox-env/bin/pip install kinescript-chatterbox
export KINE_VOICE_CMD_CHATTERBOX='["'$HOME'/chatterbox-env/bin/python", "-m", "kinescript_chatterbox"]'

The first run downloads Chatterbox Turbo (4.0 GB) and wav2vec2 (380 MB) from Hugging Face. The standard model is a separate download, made the first time a sound asks for it.

Tags

Turbo (in chatterbox-tts 0.1.7) knows 19 tags:

• sounds: [laugh], [chuckle], [sigh], [gasp], [groan], [cough], [clear throat], [sniff], [shush]
• delivery: [whispering], [crying], [angry], [happy], [sarcastic], [surprised], [fear], [dramatic], [narration], [advertisement]

A tag written the way ElevenLabs writes it, such as [laughs], [whispers] or [clears throat], is read as Turbo's [laugh], [whispering] or [clear throat], so one scene works with both. A tag Turbo does not know is left out, so it is not read aloud, and listed under ignored in what kine voice prints.

Voices and settings

speaker is a WAV clip of 5 s or more, relative to the scene file, whose voice Chatterbox copies. Without one it speaks in the model's own voice. rate is not supported; Chatterbox speaks at one speed, and kine voice lists a rate other than 1 under ignored.

options:

• model: turbo (the default) acts on tags. standard is the larger original model, which has no tags but takes exaggeration (0.5 by default; higher is more emotional) and cfg_weight (lower is slower and more deliberate).
• temperature, top_p, top_k (Turbo), min_p (standard), repetition_penalty: sampling settings.
• seed: 0 by default. The same text, voice and seed speak the same way again on the same machine.

Chatterbox puts Resemble AI's inaudible Perth watermark in everything it speaks.

Word times

Chatterbox gives no word times, so the plugin finds them: wav2vec2 (facebook/wav2vec2-base-960h) hears the letters in each 20 ms of the sound, and a forced alignment fits the text's letters to them. Numbers are aligned as they are read ("250" as two hundred fifty, "2026" as twenty twenty six, "21st" as twenty first). A tag is matched to whatever sound fills its place, so a laugh does not stretch the words around it. Word times come for English only.

Device

It runs on the GPU when the installed PyTorch was built for it, and on the CPU otherwise. CHATTERBOX_DEVICE=cpu or CHATTERBOX_DEVICE=cuda picks one.