Skip to content
Sign in
Artwork for My Weird Prompts
My Weird Prompts · October 10 · 28 min

When Your TTS Model Updates Itself

Daniel builds this show on Chatterbox, Resemble AI's open source text-to-speech model — and one day he noticed our pipeline was pulling new weights he never asked for. That sent us down a rabbit hole: what actually changes between TTS point releases, and how does a downstream project silently inherit the improvement? We dig into V3's data-only gains, a code-only fix that cut Spanish tail noise 388x, Turbo's one-step decoder, and the version number that couldn't tell two very different installs apart. Along the way: the exaggeration knob that made Chatterbox famous, the PerTh watermark on every output, and why character error rate still can't measure the thing listeners actually hear. Episode 5842 — myweirdprompts.com/n/5842 (permanent short link: myweirdprompts.com/438394)

0:00 · Intro-28:49
Play in
Transcript file

transcript

No transcript — this publisher did not publish one.

show notes

Daniel builds this show on Chatterbox, Resemble AI's open source text-to-speech model — and one day he noticed our pipeline was pulling new weights he never asked for. That sent us down a rabbit hole: what actually changes between TTS point releases, and how does a downstream project silently inherit the improvement? We dig into V3's data-only gains, a code-only fix that cut Spanish tail noise 388x, Turbo's one-step decoder, and the version number that couldn't tell two very different installs apart. Along the way: the exaggeration knob that made Chatterbox famous, the PerTh watermark on every output, and why character error rate still can't measure the thing listeners actually hear.

Episode 5842 — myweirdprompts.com/n/5842 (permanent short link: myweirdprompts.com/438394)

chapters

5 chapters

more episodes

All episodes