# Making TTS "full-cast audio dramas" — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/1e66419c470b3910

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-28 17:49 ingested 2026-10-29 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xhv6j0/making_tts_fullcast_audio_dramas/)

## retained text



```
This is some demo output from [tts-audiobook-tool](https://github.com/zeropointnine/tts-audiobook-tool) (my app), where the narrator plus three speaking characters each have their own voice.

What's maybe more interesting is that, in an optional preparation step, an LLM is used to identify the characters and their lines of dialog, which you then map to the voice clones configured in your project. This process ends up being pretty reliable: It ID'ed the dialog text from the novel "The Sun Also Rises" with over 99% accuracy using Deepseek Flash 4.1 API (though the LLM's pre-existing knowledge of that novel probably helped, too). A locally served version of, say, Qwen 3.8 Flash Next ought to do quite well, too.

The TTS model used here is Breeze 2 TTS (Sounds not bad, am I right?). The app supports a dozen or two other models, but Breeze 2 is my favorite at the moment, alongside MOSS TTS 9B.

\- [Source link](https://zeropointnine.github.io/tts-audiobook-tool/browser_player/?url=https://zeropointnine.github.io/tts-audiobook-tool-sample-output/through-the-looking-glass.abr.m4a)


```



document fetched 2026-10-29 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xhv6j0/making_tts_fullcast_audio_dramas
