Paradee-8M Squeezes Voice AI Into 8.5 MB Running 62x Faster Than Real Time
Paradee-8M v1.1 is an 8.5MB text-to-speech model distilled from Kokoro-82M that runs 62x faster than real time on a single CPU thread.
Takeaways
- Paradee-8M v1.1 released: 8.5 MB ONNX text-to-speech model distilled from Kokoro-82M.
- Runs 62x faster than real time on one CPU thread, up from 18x in v1.0.
- UTMOS 4.39 vs Kokoro's 4.52, word error rate 5.9% vs 5.7%, on 200 held-out sentences.
- New decoder head skips upsampling, builds voiced sound from pitch-aligned sine waves to kill buzz.
- Apache 2.0, trained on 12,000 WikiText-103 sentences on a single MacBook Pro.
- Code and weights available; speaks one voice (
af_heart) in American English.
Paradee-8M v1.1 packages a distilled version of Kokoro-82M into an 8.46 MB, int8-quantized ONNX model. The release generates 24 kHz English speech at a reported 62.1 times real time on one CPU thread, allowing one second of audio to render in about 16 milliseconds without a GPU.
Version 1.1 increases throughput from v1.0’s 18.5 times real time while keeping word error rate at 5.9% and moving UTMOS from 4.41 to 4.39. That combination makes the model relevant for offline applications where download size, latency, and CPU use constrain deployment.
A faster decoder removes two costly stages
Most of the speedup comes from a redesigned audio-generation head. Version 1.0 used a narrower form of Kokoro’s generator, starting with representations at 80 frames per second and progressively upsampling them to 24,000 audio samples per second. Computation increased during each stage as the sequence grew, and voiced sounds acquired a slight buzz that required a separate phase-correction filter.
Version 1.1 predicts its controls at 80 frames per second and constructs the final waveform directly. For voiced sounds, it combines sine waves at the estimated fundamental frequency and its harmonics, then learns gain values for individual frequency bands and small timing corrections. Pitch-aligned oscillators remove the need for the earlier phase-correction filter.
Breath, hiss, and consonants come from shaped random noise. One noise path pulses with the estimated pitch to model breathy voiced sounds. Another operates in three-millisecond steps, preserving sharp consonants and word onsets that a slower control rate could smear.