Orukeet Core ML preview for TapTalk

Portable Orukeet r3 Core ML model bundles for TapTalk's FluidAudio 0.15.5 runtime.

Implementation PR #6 includes the conversion code, Swift integration library, and raw validation evidence.

Downloads

• orukeet-r3-coreml-greedy.zip: recommended for ordinary greedy decoding; removes unused top-64 joint calculations.
• orukeet-r3-coreml-baseline.zip: exact Parakeet graph structure with Orukeet weights; retains top-64 outputs for language hints/reranking.
• SHA256SUMS.txt: SHA-256 archive checksums. Each archive also contains a per-file bundle.json manifest, source receipts, attribution, and CC BY-SA 4.0 weight license.

The archives include four .mlpackage components and the 8192-token vocabulary. Compile packages on the destination Mac during installation; machine-specific compiled caches are excluded.

Swift integration and conversion guide · Benchmark and validation evidence

Measured results and limits

On an M5 Max, baseline latency was within 0.3% of Parakeet and the greedy profile reduced batch latency by 9.5%. Greedy and baseline text matched on all 128 recordings in the eight-language regression sample. Core ML WER was 7.64%, versus 8.55% for Parakeet and 7.29% for the uncompressed Orukeet source.

This is a preview. Other Macs and OS versions are unmeasured. Default sliding-window TDT streaming still buffers about 13 seconds before first text; this is separate from Parakeet EOU 120M. The existing stable Metal release remains unchanged.

Orukeet 0.1.1 — optimized Metal runtime

Orukeet 0.1.1 adds prebuilt optimized Metal inference on Apple silicon. The Python installer downloads the SDK from this release and verifies its SHA-256. Model weights remain the existing r3 v0.1.0 artifacts.

With Python 3.12+ in an activated virtual environment, after this release is published:

python -m pip install --upgrade https://github.com/Oruk-AI/orukeet/releases/download/v0.1.1/orukeet-0.1.1-py3-none-any.whl
orukeet install --device auto --cache ./orukeet-cache --output installation.json

Use the installation receipt to transcribe locally:

import json
from pathlib import Path
from orukeet import Orukeet

config = json.loads(Path("installation.json").read_text(encoding="utf-8-sig"))
with Orukeet(config["model"], config["runtime"], device=config["device"]) as asr:
    print(asr.transcribe("recording.wav")["text"])

Automatic selection uses the optimized Metal SDK on Apple silicon. CPU, CUDA and Vulkan entries retain their existing SDKs. Existing users should upgrade the package and regenerate installation.json to select the new runtime. This GitHub release does not publish a package to PyPI.

The SDK is built from the pinned sources and Metal/attention patches in Orukeet's runtime/ directory. It retains NeMo's 0.1.0 SDK version and v1 C ABI. Prebuilt installation requires no source checkout or compiler. Build and packaging instructions are in runtime/README.md and runtime/RELEASE.md.

Candidate validation on Apple M4 Pro / macOS 26.5:

• Two packaging runs produced identical SDK archives of 1,628,856 bytes.
• The archive passed installation, automatic Metal selection, offline cache reuse, and JFK transcription with worker reuse.
• All 48 recognitions across 24 test clips matched the reference transcripts.
• The rebuilt 0.1.1 wheel installed in an isolated environment, installed the candidate SDK and transcribed successfully.
• Package tests: 36 passed, one model-dependent test skipped.

Before publication, installer verification redirects only the SDK URL to the exact local candidate archive and uses existing verified model weights. Public download verification must follow publication. The attached metal-validation.json records this scope; these checks do not establish new latency or full-model numerical measurements.

The SDK contains its build provenance and dependency notices. metal-runtime.json is the pinned installer entry; SHA256SUMS covers the attached artifacts.

Orukeet v0.1.0

Orukeet is a 25-language ASR model built from Parakeet TDT 0.6B v3 with 12,288 fitted, frozen Gabor kernels.

This release contains the final r3 checkpoint in NeMo, Q8 and F16 formats, with pinned hashes. The inference package includes persistent workers and native Metal, CUDA, Vulkan and CPU runtimes.

Orukeet has lower WER on 61 of 74 tested splits. FLEURS pooled WER is 9.85% versus 11.01% for Parakeet. The model card and report include every paired score and the evaluation protocol.

• Weights and model card
• Run Orukeet
• All paired benchmark scores
• OpenWhispr integration

Code: MIT. Weights and fitted kernels: CC BY-SA 4.0. Numerical evaluation records: CC BY 4.0. NVIDIA's foundation attribution is retained.

The report PDF, BibTeX and independently verified arXiv source package are attached. The paper has not been submitted to arXiv.