Neural ExpeditionSubscribe
TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #88

Clone a voice from ten seconds and make it laugh on cue

Longtime readers met Chatterbox back in March: Resemble AI's open voice model that clones and emotes in 23 languages. It had one weakness shared by every TTS that sounds human: too slow for a live conversation. Turbo is the same team's answer, the model distilled down until it fits a real-time voice agent, published under MIT like its sibling.

MODELchatterbox-turbo
PUBLISHEDAugust 10, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYAUDIO

Field notes

01What it does

Chatterbox-Turbo is a 350 million parameter text-to-speech model that clones a voice from a short reference clip, no training involved. You give it roughly ten seconds of someone speaking and a line of text, and it reads the text in that voice. The party trick is native paralinguistic tags: write [laugh], [chuckle], or [cough] inside your sentence and the voice performs it in place, which is what makes agent conversations stop sounding like dictation.

The Turbo part is real engineering, not branding: the decoder that turns speech tokens into audio used to take ten steps and now takes one, so the model needs less compute and less video memory than the earlier 500 million parameter Chatterbox models. It was built for voice agents that must answer in real time, and it holds up for narration too.

02How to try it

The official demo Space is the fastest route: type a sentence, drop a tag or two in it, upload a short voice clip and listen. There is also a samples page if you want to hear it before trying. Locally, pip install chatterbox-tts and about six lines of Python load the model and save a wav file; the README shows the exact snippet, reference clip included.

03Caveat

Turbo speaks English only: the multilingual sibling covers 23 languages but is a step slower and heavier. Cloning quality follows the reference clip, so a noisy or accented sample carries its flaws into every sentence. Every output also carries Resemble's PerTh audio watermark, inaudible but detectable, which is arguably a feature. And the obvious one: clone voices you have permission to clone.

04What you can do with it

  • Give a phone or support agent a voice that chuckles and hesitates naturally.
  • Narrate videos or a podcast intro in your own cloned voice.
  • Voice game characters with performed laughs and coughs instead of flat lines.
  • Read your articles or newsletters aloud in a consistent branded voice.
  • Prototype an assistant's voice on a laptop before paying for a studio.

Try the demo

View model page

Read this issue on neuralexpedition.com →