kokoro-82m-ane-coreml

Kokoro 82M — Apple Neural Engine Core ML chain

Seven-stage Core ML conversion of hexgrad/Kokoro-82M (Apache-2.0) for on-device text-to-speech entirely on Apple's Neural Engine, adapted from laishere/kokoro-coreml (Apache-2.0).

Consumed by Vox as the kokoro-82m-ane TTS engine. Benchmark (M1 Pro): 19–22.6× realtime flat, 75–850 ms time-to-first-chunk.

Layout

• KokoroAlbert.mlmodelc, KokoroPostAlbert.mlmodelc, KokoroAlignment.mlmodelc, KokoroProsody.mlmodelc, KokoroNoise.mlmodelc, KokoroVocoder.mlmodelc, KokoroTail.mlmodelc — compiled Core ML stages. All seven must load with .cpuAndNeuralEngine ("strict" placement): mixed placement costs 2–4×, and Prosody on .all segfaults on macOS 27 (mps.expand_dims invalid axis). Static shapes: ALBERT ≤ 510 phonemes, vocoder ≤ 2000 frames.
• vocab.json — Kokoro phoneme→token map.
• voices/*.bin — raw fp32 [510, 256] voice embeddings (row = token count − 1; first 128 dims timbre, last 128 style). Converted from the upstream voices/*.pt tensors; the conversion is byte-identical to the iOSDemo's reference af_heart.bin.

Noise-stage fix (2026-10-20)

KokoroNoise.mlmodelc/model.mil is patched (weights unchanged); noise-phase-fix.json marks the fixed revision. The conversion spelled atan2 as atan(im/re) plus quadrant terms with no case for im == 0, re < 0, so the DC and Nyquist STFT bins of the harmonic source got phase 0 where torch.angle gives +π, and it replaced the source's Gaussian noise (noise_amp × N(0, 1)) with a constant. Together these lifted 5–10 kHz by up to 2.8 dB against the MLX port. The fix makes the +π branch take im >= 0 and restores the draw with random_normal (seed −1: the same draw on every load, so output stays deterministic). Measured in Vox's VoiceBench (16 cells × 3 takes, M1 Pro): the long-term spectrum is within 0.1 dB of MLX Kokoro from 160 Hz to 10 kHz (was 2.8 dB), centroid +2 Hz (was +195 Hz). The previous revision is e16ab079.

Noise stage v2 (2026-10-20)

KokoroNoise.mlmodelc is rebuilt from kokoro-v1_0.safetensors with the phase fix and the noise draw built in; noise-stage-v2.json marks this revision. The harmonic source and the STFT stay fp32; the noise branch (convs, Snake, AdaIN) is fp16, and the weights are not palettized. The stage runs on the CPU in about half the time (33 ms against 72–76 ms for a 370-frame input, M1 Pro). VoiceBench: every quality budget passes, as with the patched stage; time to first audio p50 dropped 35% in a paired A/B. The previous revision is 3ddd693.

Voices (English launch set)

American: af_alloy, af_aoede, af_bella, af_heart, af_jessica, af_kore, af_nicole, af_nova, af_river, af_sarah, af_sky, am_adam, am_echo, am_eric, am_fenrir, am_liam, am_michael, am_onyx, am_puck, am_santa. British: bf_alice, bf_emma, bf_isabella, bf_lily, bm_daniel, bm_fable, bm_george, bm_lewis.

License

Model weights and voices: Apache-2.0 (hexgrad/Kokoro-82M). Core ML conversion: Apache-2.0 (laishere/kokoro-coreml, © 2026 laishere). Voice pack conversion scripted in Vox: scripts/generate-ane-voice-packs.py.