robbiemu/paper2-kokoro-coreml-components

Kokoro Core ML components

Twelve downloadable Core ML packages: FP32 and FP16 duration heads, plus conditioning and generator pairs for 115, 127, 142, 153, and 190 frames.

Model identity and compatibility

• Upstream model: hexgrad/Kokoro-82M. The study checkpoint is pinned to f3ff3571791e39611d31c381e3a41a3af07b4987.
• Derivation: Core ML conversion of selected duration and decoder components, with precision and graph-layout alternatives.
• Language scope: English is the retained study scope; this release does not establish multilingual compatibility.
• Interface: The duration heads take predictor states with shape [1,128,512]. Each conditioning/generator pair has one fixed alignment: 115, 127, 142, 153, or 190 frames; the generator returns magnitude and phase arrays. Host-side alignment, pitch/noise construction and waveform reconstruction remain outside these packages.
• Validation scope: Duration comparisons cover 24 selected requests. The historical saved-feature waveform composition covers one fixed input with separate hardware replay. The live E3 study covers five development requests, one voice, and five shapes: four meet the frozen Mac CPU criteria and ordinary_006 fails the log-magnitude MAE limit. This is not a general text-to-audio quality evaluation.
• Data role: Conversion and replay use retained upstream-model inputs and outputs; no new speech-model pretraining is claimed.

The base-model link records component lineage. Conversion, graph adjustment and fitting are described separately above. Core ML input/output dtypes do not by themselves specify internal compute precision.

The documented Python Core ML loading path requires a compatible macOS Core ML runtime. Recorded Core ML Tools versions: 9.0. Exact specification versions, conversion dates, source-framework versions and I/O dtypes are preserved in coreml_specs.json; current cross-version runtime compatibility has not been retested.

release_metadata.json preserves this scope as machine-readable data. The private evidence dataset is linked for experiment records and is not labeled as a training dataset.

File and tensor details

This release uses Core ML packages. It contains no Safetensors weights, so the Hub's automatic Safetensors panel does not apply.

Hugging Face's Safetensors metadata parser reads tensor headers from Safetensors files; README metadata alone cannot supply that panel.

Format	Packages or files	Stored size
Core ML package	12	509.79 MiB

Browse the per-file details, or inspect tensor names, shapes and storage dtypes. These sizes include separate alternatives; they are not a combined model parameter count.

Package	Stored size	Input → output dtypes	Specification
models/conditioning-115frames.mlpackage	64.15 MiB	FLOAT32 → FLOAT32	parseable
models/conditioning-127frames.mlpackage	64.15 MiB	FLOAT32 → FLOAT32	parseable
models/conditioning-142frames.mlpackage	64.15 MiB	FLOAT32 → FLOAT32	parseable
models/conditioning-153frames.mlpackage	64.15 MiB	FLOAT32 → FLOAT32	parseable
models/conditioning-190frames.mlpackage	64.15 MiB	FLOAT32 → FLOAT32	parseable
models/duration-fp16.mlpackage	54,382 B	FLOAT32 → FLOAT32	parseable
models/duration-fp32.mlpackage	105,017 B	FLOAT32 → FLOAT32	parseable
models/generator-115frames.mlpackage	37.78 MiB	FLOAT32 → FLOAT32	parseable
models/generator-127frames.mlpackage	37.78 MiB	FLOAT32 → FLOAT32	parseable
models/generator-142frames.mlpackage	37.78 MiB	FLOAT32 → FLOAT32	parseable
models/generator-153frames.mlpackage	37.78 MiB	FLOAT32 → FLOAT32	parseable
models/generator-190frames.mlpackage	37.78 MiB	FLOAT32 → FLOAT32	parseable

Core ML input/output dtypes describe the interface, not internal compute or weight precision.

Model files

Model / configuration	Files	Role
duration-fp32	models/duration-fp32.mlpackage	Retained CPU FP32 duration head
duration-fp16	models/duration-fp16.mlpackage	FP16 duration comparison
conditioning-115frames	models/conditioning-115frames.mlpackage	Conditioning suffix, fixed 115 frames
generator-115frames	models/generator-115frames.mlpackage	Converted 4-D generator, fixed 115 frames
conditioning-127frames	models/conditioning-127frames.mlpackage	E3 static 127-frame conditioning
generator-127frames	models/generator-127frames.mlpackage	E3 static 127-frame generator
conditioning-142frames	models/conditioning-142frames.mlpackage	E3 static 142-frame conditioning
generator-142frames	models/generator-142frames.mlpackage	E3 static 142-frame generator
conditioning-153frames	models/conditioning-153frames.mlpackage	E3 static 153-frame conditioning
generator-153frames	models/generator-153frames.mlpackage	E3 static 153-frame generator
conditioning-190frames	models/conditioning-190frames.mlpackage	E3 static 190-frame conditioning
generator-190frames	models/generator-190frames.mlpackage	E3 static 190-frame generator

Download and load

from huggingface_hub import snapshot_download
from pathlib import Path
import coremltools as ct

root = Path(snapshot_download("robbiemu/paper2-kokoro-coreml-components"))
package = root / "models/duration-fp32.mlpackage"
spec = ct.utils.load_spec(str(package))
print(spec.description.input)
model = ct.models.MLModel(str(package), compute_units=ct.ComputeUnit.CPU_ONLY)
# Supply the named arrays and shapes listed in coreml_specs.json:
# output = model.predict(inputs)

The example selects CPU execution explicitly. Use the study’s declared compute policy when reproducing a comparison. coreml_specs.json lists exact input/output names, shapes and conversion metadata. manifest.json records file sizes, SHA-256 hashes and provenance paths. python restore.py --verify checks all model files.

Scope and result

The retained composition uses the CPU FP32 duration policy and preserves source pitch, noise and harmonic construction around the converted suffix. The historical saved-feature waveform acceptance covers one fixed input with separate hardware replay. The later live study retains a numerical failure on one of five requests. These components require the host orchestration and remaining upstream Kokoro operations; they are not a complete text-to-audio checkpoint.

These model files were published from retained study artifacts. Their integrity and Core ML interfaces were checked; publication does not constitute a new inference or deployment validation. Experimental evidence, diagnostic probes, transfer intermediates and optimization history are held in separate private repositories.

Provenance and license

Upstream notices/model cards are retained under licenses/; the applicable model license is in LICENSE. The private evidence dataset records historical source revisions, execution policies and results.

Study evidence

The additional conditioning/generator pairs cover 127, 142, 153 and 190 frames. Package interfaces, sizes and hashes are listed in the model metadata above. The study report and study records are retained in the private evidence dataset. The complete E3 archive also contains the source-equivalence records, captured arrays, code and runtime receipts. Access to these records requires dataset permission. Public model files do not provide anonymous access to the complete scientific evidence.

This repository follows the other component releases: model packages, interfaces, manifests, licenses and model-card documentation. The study's scientific records are maintained in the evidence dataset. No model bytes or scientific results changed.