Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model
arXiv cs.AI · 6 Oct 2026, 03:56 AEST
Paradee is an 8.07M-parameter text-to-speech model distilled from the Kokoro-82M model. It operates with significantly fewer parameters and reduced compute requirements while maintaining high performance.
Key facts
- We distill Kokoro-82M, a widely used open text-to-speech model with 54 voices, into Paradee, an 8.07M-parameter model that speaks one of them.
- Paradee keeps Kokoro's architecture with much narrower layers, and each of its two halves is trained separately against the frozen teacher.
- It has 10x fewer parameters and needs 15x less compute.
- Stored in int8, Paradee is 8.5 MB, runs 25x faster than real time on one CPU thread, and scores 4.41 on UTMOS against the teacher's 4.52.
Read aloud
Paradee is an eight.07M-parameter text-to-speech model distilled from the Kokoro-82M model. It operates with significantly fewer parameters and reduced compute requirements while maintaining high performance. We distill Kokoro-82M, a widely used open text-to-speech model with fifty four voices, into Paradee, an eight.07M-parameter model that speaks one of them. Paradee keeps Kokoro's architecture with much narrower layers, and each of its two halves is trained separately against the frozen teacher. It has 10x fewer parameters and needs 15x less compute. Stored in int8, Paradee is eight point five MB, runs 25x faster than real time on one CPU thread, and scores four point four one on UTMOS against the teacher's four point five two.
Sources in this cluster
- Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model
arXiv cs.AI · 6 Oct 2026, 03:56 AEST