qmohits/kokoro-82m-coreml

Kokoro-82M Core ML

Fixed-shape Core ML packages converted from hexgrad/Kokoro-82M snapshot f3ff3571. They were graded on a physical iPhone 12 (A14, 4 GB).

The conversion source is the PyTorch checkpoint. The public Kokoro ONNX model is the listening reference. Kokoro weights stay Apache-2.0. This repository does not include the original checkpoint, the ONNX file, or a phonemizer.

Conversion scripts, the iPhone benchmark, frozen inputs, and the raw result JSON are in the GitHub repository:

github.com/qmohitsingh/kokoro-82m-coreml

Graded on iPhone 12

Voice af_heart, speed 1.0, iOS 18, Core ML compute units .all. Two warm-ups were discarded, then the median of 10 runs. The phonemizer is outside the timer: token ids and a style row go in, and a 24 kHz waveform comes out.

Passage	Median RTF	Samples matched	Thermal state
3 s	0.193	yes	serious
7 s	0.202	yes	serious
15 s	0.220	yes	serious
30 s	0.252	yes	serious

RTF is wall seconds divided by audio seconds. The 30 second passage is two model calls joined together. Its slowest scored run was 0.300.

A blind comparison against the ONNX reference was four ties. The reference was not preferred on any passage. A continuous Core ML loop then ran for 3,606 seconds and 316 loops without a crash.

Packages

Each passage has five fp32 packages. The same weights are stored again for every length, so a download of all 25 packages is about 1.5 GB.

Passage	Tokens	Frames	Audio	Package suffix
3 s	56	146	3.650 s	t56 and f146
7 s	110	275	6.875 s	t110 and f275
15 s	237	567	14.175 s	t237 and f567
30 s, first call	260	614	15.350 s	t260 and f614
30 s, second call	280	633	15.825 s	t280 and f633

The five stages are:

Package	Inputs	Outputs
kokoro_text_t*_fp32	input_ids	encoded
kokoro_prosody_text_t*_fp32	input_ids, encoded, style	duration_hidden, duration_logits, text_features
kokoro_pitch_energy_f*_fp32	aligned, style	pitch, energy
kokoro_decoder_body_f*_fp32	text, pitch, energy, style	features
kokoro_generator_f*_fp32	features, style, spectrum	magnitude, phase

style is the 256-value voice row. Pitch and energy use the prosody half. The decoder body and generator use the timbre half. A package only accepts the token count or frame count in its name.

Swift, in the iPhone benchmark, repeats each token by its rounded duration, builds the harmonic source, and turns magnitude and phase back into audio. Those steps are not inside these packages.

Download one passage

The 3 second set is the five packages ending in t56 or f146:

hf download qmohits/kokoro-82m-coreml \
  --include "kokoro_text_t56_fp32.mlpackage/**" \
  --include "kokoro_prosody_text_t56_fp32.mlpackage/**" \
  --include "kokoro_pitch_energy_f146_fp32.mlpackage/**" \
  --include "kokoro_decoder_body_f146_fp32.mlpackage/**" \
  --include "kokoro_generator_f146_fp32.mlpackage/**"

Limits

These graphs are fixed at conversion time. A different sentence length needs a new pair of shapes. One call cannot cover the full 30 second passage, because the phoneme cap is 510. The benchmark app is about 1.9 GB for the same reason the packages repeat their weights.