robbiemu/paper2-apple-tts-fitted-corrections

Apple TTS fitted corrections and Core ML codecs

Six downloadable frozen FP16 weight arrays in three configurations, plus two static Core ML codec model packages.

Model identity and compatibility

• Upstream model: BreezeBlue/Breeze-TTS-2. The study checkpoint is pinned to 799624c0b4a1daa8db6d28bbd9850043c0270734.
• Derivation: Fitted replacement weights for two depth-transformer key projections, an uncorrected FP16 control, and separate Core ML codec conversions.
• Language scope: English and Mandarin are the retained bilingual confirmation scope, matching the upstream language declaration.
• Interface: Replacement arrays are frozen, merged FP16 executing weights, loaded with NumPy and installed in the PyTorch/MPS generator. They are not additive deltas. The two Core ML codecs take integer codes with shape [1,16,38] and return 72,960 waveform samples. The full generator remains a separate dependency.
• Validation scope: Each language supplies 24 fitting, 8 selection and 8 held-out confirmation requests, with four earlier requests retained as guards. Neither fitted alternative passed the frozen bilingual acceptance rules; the incumbent remained selected.
• Data role: The fitted parameters use saved internal outputs from newly written synthetic utterances. Confirmation waveform comparisons are evaluation, not pretraining. The static codec packages are a separate conversion experiment.

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 and NumPy parameter arrays. 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	2	653.07 MiB
npy	6	3.00 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/codec-fp16-38frames.mlpackage	217.80 MiB	INT32 → FLOAT16	parseable
models/codec-fp32-38frames.mlpackage	435.27 MiB	INT32 → FLOAT32	parseable

File	Stored size	Tensors	Tensor elements	Storage dtypes
corrections/axis_1_minus/depth-key-0.fp16.npy	524,416 B	1	262,144	float16
corrections/axis_1_minus/depth-key-6.fp16.npy	524,416 B	1	262,144	float16
corrections/joint_ridge_2/depth-key-0.fp16.npy	524,416 B	1	262,144	float16
corrections/joint_ridge_2/depth-key-6.fp16.npy	524,416 B	1	262,144	float16
corrections/uncorrected/depth-key-0.fp16.npy	524,416 B	1	262,144	float16
corrections/uncorrected/depth-key-6.fp16.npy	524,416 B	1	262,144	float16

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

Model files

Model / configuration	Files	Role
codec-fp16-38frames	models/codec-fp16-38frames.mlpackage	Static FP16 codec, 38 frames
codec-fp32-38frames	models/codec-fp32-38frames.mlpackage	Static FP32 codec, 38 frames
axis_1_minus	corrections/axis_1_minus	Fitted replacement weights; rejected by confirmation gate
joint_ridge_2	corrections/joint_ridge_2	Fitted replacement weights; rejected by confirmation gate
uncorrected	corrections/uncorrected	Uncorrected demotion control

Download and load

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

root = Path(snapshot_download("robbiemu/paper2-apple-tts-fitted-corrections"))
package = root / "models/codec-fp16-38frames.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.

Load an array with numpy.load(path, allow_pickle=False) and replace the corresponding projection’s executing weight. correction_config.json identifies the target names and hashes. Preserve all other incumbent remedies; do not stack the correction onto another fitted version.

Scope and result

Derived from Breeze TTS 2 by BreezeBlue and licensed for research and non-commercial use only. The generator arrays are replacement weights for two depth-transformer key projections, already merged and rounded once to FP16. They are not deltas to add or full generator checkpoints. Neither fitted alternative passed the frozen bilingual acceptance gate; the incumbent was retained. The codec exports are separate, fixed-38-frame components.

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.