robbiemu/paper2-dia1-dac-coreml-decoders

Dia1 / DAC Core ML decoders

Twelve downloadable DAC decoder model packages: four precision alternatives and eight fitted or graph-adjusted candidates.

Model identity and compatibility

• Upstream model: descript/dac_44khz. The current Hub revision is not asserted to be the historical download revision; retained parameter hashes identify the study artifacts.
• Derivation: Core ML DAC decoder conversions plus graph adjustments and fitted decoder corrections. Dia1 uses this codec; no Dia1 text-generator weights are included.
• Language scope: These are acoustic decoders operating on latent features. No text-language capability is assigned to the decoder release.
• Interface: Every public package takes continuous quantized latent features with shape [1,1024,285] and returns a waveform with shape [1,1,145920]. Packages have fixed shapes; FP16, FP32 and mixed-compute variants can still expose FP32 input/output arrays.
• Validation scope: The four precision alternatives were compared on two retained inputs. Adaptive candidates have separate development evidence; the channel-affine candidate worsened complete-waveform error. Candidate inclusion does not imply deployment acceptance.
• Data role: Fitting and evaluation data differ across candidates. The source attribution file and per-candidate evidence receipts distinguish calibration, development and held-out data; they do not imply that every listed corpus trained every variant.

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	1,459.25 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/baseline-fp32.mlpackage	206.46 MiB	FLOAT32 → FLOAT32	parseable
models/baseline-fp16.mlpackage	103.28 MiB	FLOAT32 → FLOAT32	parseable
models/mixed-fp32-convolution.mlpackage	206.44 MiB	FLOAT32 → FLOAT32	parseable
models/mixed-fp32-snake-residual.mlpackage	103.33 MiB	FLOAT32 → FLOAT32	parseable
models/candidate1-affine.mlpackage	103.28 MiB	FLOAT32 → FLOAT32	parseable
models/candidate2-response.mlpackage	103.28 MiB	FLOAT32 → FLOAT32	parseable
models/candidate3-roi.mlpackage	103.28 MiB	FLOAT32 → FLOAT32	parseable
models/candidate4-spectral.mlpackage	103.28 MiB	FLOAT32 → FLOAT32	parseable
models/candidate4-shape.mlpackage	103.28 MiB	FLOAT32 → FLOAT32	parseable
models/candidate5-polyphase.mlpackage	107.78 MiB	FLOAT32 → FLOAT32	parseable
models/candidate5-channel-affine.mlpackage	107.78 MiB	FLOAT32 → FLOAT32	parseable
models/candidate6-response.mlpackage	107.78 MiB	FLOAT32 → FLOAT32	parseable

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

Model files

Model / configuration	Files	Role
baseline-fp32	models/baseline-fp32.mlpackage	Preferred tested CPU alternative
baseline-fp16	models/baseline-fp16.mlpackage	FP16 comparison
mixed-fp32-convolution	models/mixed-fp32-convolution.mlpackage	FP32 convolutions / FP16 periodic operations
mixed-fp32-snake-residual	models/mixed-fp32-snake-residual.mlpackage	FP32 Snake/residual operations comparison
candidate1-affine	models/candidate1-affine.mlpackage	Shared affine fitted decoder
candidate2-response	models/candidate2-response.mlpackage	Response-surface fitted decoder
candidate3-roi	models/candidate3-roi.mlpackage	M4 region-of-interest correction
candidate4-spectral	models/candidate4-spectral.mlpackage	C4 in Appendix C; baseline and parent of C5
candidate4-shape	models/candidate4-shape.mlpackage	M4 shape-response correction
candidate5-polyphase	models/candidate5-polyphase.mlpackage	Polyphase decoder
candidate5-channel-affine	models/candidate5-channel-affine.mlpackage	Fitted variant; worse complete-waveform error
candidate6-response	models/candidate6-response.mlpackage	Partial development-set improvement

Appendix C candidate identities

The M4 comparisons in Appendix C use candidate4-spectral as C4. candidate5-polyphase (C5) rewrites its upsampling operation while retaining the existing corrections. candidate5-channel-affine and candidate6-response are separate modifications of C5. The separately released candidate4-shape is a different C4 variant.

candidate_lineage.json records these relationships, exact package directory hashes and source-configuration hashes. manifest.json records every package file hash. The comparisons reuse 32 development inputs; independent acceptance and baseline-only calibration remain unestablished.

Download and load

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

root = Path(snapshot_download("robbiemu/paper2-dia1-dac-coreml-decoders"))
package = root / "models/baseline-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 fixed precision comparison preferred the all-FP32 CPU decoder on two retained inputs: it had both lower error and shorter call time than the mixed-convolution alternative. The adaptive candidates are a separate study. Candidate 5 channel-affine fitting worsened waveform error; Candidate 6 showed only a partial, metric-dependent development result. Inclusion preserves those models for research, not a claim that every candidate is a deployment improvement. The historical conversion frontend warning remains applicable.

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.