Compressed Kokoro: listen and compare
The same sentence from Kokoro-82M and four smaller decoders distilled from it. Every version shares the same phonemes, timing and prosody; only the waveform decoder differs. Switch between versions mid-sentence: playback continues from the same moment, the best way to hear small differences.
1 · Pick a sentence
2 · Listen and switch
Keys: 1–5 switch version · space play/pause · ←/→ skip 2 s.
Headphones help. In blind mode the order is shuffled and labels stay hidden until you reveal them.
The variants
| version | whole model | decoder speed | voiced artifact rate | blind test vs Kokoro |
|---|
Decoder speed is relative to Kokoro's decoder on an Apple M4 GPU (PyTorch). The artifact rate is the share of voiced frames with a narrowband 2–3 kHz excess over Kokoro, a faint "hoarse whistle" (lower is better). Blind tests: one listener, loudness-matched, with hidden repeats. Samples are sentences from public-domain books not used in training, 24 kHz.