Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model
Paradee distills Kokoro-82M into an 8M single-voice TTS model scoring 4.41 UTMOS.
Paradee distills Kokoro-82M, an open 54-voice text-to-speech model, into an 8.07M-parameter single-voice model with roughly 10x fewer parameters and 15x less compute. A small text model and decoder are trained separately against the frozen teacher, then joined and quantized to int8. Stored in int8, Paradee is 8.5 MB, runs 25x faster than real time on one CPU thread, and scores 4.41 UTMOS versus the teacher's 4.52. A training-free phase-locking filter removes most residual buzz between 2 and 8 kHz.
- 8.07M parameters versus Kokoro-82M: about 10x smaller
- Text and decoder halves train separately against a frozen teacher
- Int8 model is 8.5 MB and 25x faster than real time on one CPU thread
- UTMOS is 4.41 versus the teacher's 4.52
- AI models
- ParadeeKokoro-82M
From the original article
1 min read at arxiv.orgWe 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.
By Sahil Mahendrakar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.