Voice lab · Chatterbox Turbo

A performed read

One sentence, three takes. The first two come from the same model and the same voice — the only difference is a tag in the text. The third is the voice the harness speaks with today.

Chatterbox Turbo · with the tag · 4.8 s

Transcript: I could not believe it. [laugh] That was the best part of the whole trip.

The tag is not spoken. It becomes a laugh: this take is 1.04 s longer than the same text below, and two noise bursts appear in the gap between the sentences. Measured, not assumed.

Chatterbox Turbo · same words, no tag · 3.8 s

Transcript: I could not believe it. That was the best part of the whole trip.

The control. Play these two back to back — the difference is the tag and nothing else.

Kokoro af_bella · the reply voice today · 4.3 s

Transcript: I could not believe it. That was the best part of the whole trip.

Same words, no tag — this is what the assistant actually speaks with, for comparison.

What this is

Model
ResembleAI/chatterbox-turbo, MIT, 350M, run through chatterbox-tts 0.1.7
Where it runs
fleet3, in its own environment (C:\Users\jf803\chatterbox: Python 3.13.5, torch 2.6.0 CPU). It shares nothing with the Kokoro voice stack.
The voice
Cloned from an 11.7 s clean reference take of the same speaker, made for this test with his consent.
Output
24 kHz mono WAV
Measured cost
28.9 s to load the model; about 16× slower than real time on this CPU laptop — the 4.8 s take took 104 s to render. This is the real limit: a five-minute read would take about 80 minutes.
Tags, measured
[laugh] works: +1.04 s and two audible bursts between phrases. [chuckle] on a short line did not change the read here (3.04 s against 3.17 s untagged), so do not trust every tag name equally. Speech recognition hears the same words with or without a tag — the tag never leaks into the transcript.
What it is not
Not the reply voice. The assistant still speaks Kokoro af_bella; Turbo is not wired into any tool.
Limits
CPU only, English only, one short line at a time.
Correction, 2026-10-26. The first version of this page played a clip that read the file paths aloud (“ref audio take…”): a bad generation, kept honest here rather than quietly swapped. It has been replaced with the clip above, and the tag is now checked against a control instead of a single take.