coder543/parakeet-v2-coreai

Parakeet TDT 0.6B V2 — Core AI

Apple Core AI conversion of NVIDIA Parakeet V2, pinned to ae9ad07059c7c739ffaf932226a8fe64ae2620b0. Converted weights retain the upstream Creative Commons Attribution 4.0 license. See LICENSE and NOTICE. This conversion is not endorsed by NVIDIA.

Choose one self-contained bundle:

• fast/: full finite Fourier relative attention, 192 encoder positions, fixed 15-second chunks, up to 128 independent chunks in the FP16 ANE decoder.
• quality/: full sinusoidal relative attention, 3,776 encoder positions, exact tiled convolution subsampling, fixed 300-second chunks and an FP32 CPU decoder batching up to four independent chunks.

Both use W8A16 encoder weights, retaining selected sensitive projections in FP16, and four concurrent encoder requests. Neither truncates attention within its chunk or drops input audio. runtime.json records the qualified scheduling policy. The two bundles trade speed for longer context; the labels describe the measured JFK result rather than a universal quality ranking.

Requires physical Apple silicon on macOS 27 or iOS 27 and a model-specific host runtime. Graphs accept model features and recurrent states, not audio files. The host must implement the frontend, greedy TDT loop and tokenizer described by metadata/sidecars. Weights and recurrent states are shared across static functions; recurrent histories remain independent between utterances. Vocabulary: 1024 nonblank tokens, blank ID 1024, durations 0/1/2/3/4 and at most ten symbols per frame. For the larger V3 vocabulary the batched graph returns (partition, local) in two FP16 channels; reconstruct a token as partition * 2048 + local. Global IDs above 2,048 must not be transported as a single FP16 value. The scalar FP32 decoder returns a global token ID.

Measured performance

M3 MacBook Air (16 GB), macOS 27 build 26A428. Medians after warmup; frontend, encoder and decoding are included, while preparation, file I/O and chunk planning are excluded. The audio is JFK's “We choose to go to the Moon” speech: a 20-second excerpt and the 18-minute 15-second recording. The quality short-clip test intentionally uses the unmodified five-minute shape.

Bundle	Audio duration	Transcription	Audio / elapsed
fast	20 seconds	0.150 s	133.4×
fast	18 min 15 s	1.876 s	584.0×
quality	20 seconds	1.336 s	15.0×
quality	18 min 15 s	7.044 s	155.5×

Whisper-normalized WER against the supplied long-recording reference:

Bundle	Word errors	WER
fast	79/2220	3.56%
quality	44/2220	1.98%

This is one English recording, not a general or multilingual accuracy evaluation. Native quantization and FP16 decoder operation order can change token decisions. The reported full-recording runs retain all audio, with 74 fast chunks or four quality chunks. These are distinct contexts and execution policies, not a controlled kernel-only comparison.

Device specialization and loading

Source .aimodel files specialize on the device. On the same Mac, preparation of these published source assets took:

Bundle	First observed preparation with specialization	Subsequent cached preparation
fast	33 s	0.060 s
quality	10 min 7 s	0.037 s

Preparation includes model loading and compilation requested by Core AI, and excludes warmup and transcription. These are observed cache histories, not guaranteed fresh-install times: the underlying ANE cache state is not fully observable, and OS/device/application changes can require specialization again.

AoT compiled artifacts are omitted because no significant load-time benefit has been demonstrated. Authoring debug locations were removed while preserving graph signatures and operation counts. SHA256.json lists distributed payload checksums.