Paradee — Kokoro TTS at 1/10th size
Kokoro TTS, but 10x smaller. Paradee (Hugging Apps; Space live at capture time) distills Kokoro-82M into an 8M-parameter TTS: just 9 MB of weights that “speaks faster than real time on a single CPU thread and runs on your toaster.”
- Distillation target: keep Kokoro’s voice quality at 1/10th the parameters and ~1/50th the artifact size
- Single CPU thread, faster-than-realtime — no GPU, no batching, no server
- The toaster framing: the extreme end of the Small language models thesis for speech — when the model is 9 MB, TTS becomes a library, not a service
Announcement: HuggingApps tweet (44 likes; captured as clippings/x-2107890821219619294.md with demo video link).