coder543/cohere-transcribe-coreai

Cohere Transcribe — Core AI

Apple Core AI conversion of Cohere Transcribe, pinned to revision 76b8b23e8607f35f0265a23d481b338fb0e26aea. The upstream model and converted weights use Apache License 2.0.

Requires physical Apple silicon with macOS 27 or iOS 27 and a model-specific host runtime. These graphs do not accept raw audio directly. The host provides the log-mel frontend, language prompt, token embeddings, autoregressive cache management and tokenization. See metadata.json, decoder-graph.json and prompts.json for graph capacities and prompts. The encoder retains full bidirectional context within each chunk; this is offline transcription.

The 48-layer encoder uses W8A16 with selected convolution output and relative position projections retained in FP16. Four encoder stages use the complete Fourier relative-position formulation. The eight-layer decoder uses W8A16, shared batch sizes 1/8 and cache capacities 64/128/256/512/1024. It retains all previous states within a chunk. Host embedding tables use FP32.

Both encoder and decoder prefer ANE. The qualified runtime uses two encoder requests in flight, up to eight independently decoded chunks, reused output buffers and quiet-boundary chunks of at most 35 seconds. Every input sample is retained; no words or attention positions are intentionally dropped.

Measured performance

M3 MacBook Air (16 GB), macOS 27 build 26A428; median warm transcription, including frontend, encoder and decoder, excluding preparation, audio file I/O and chunk planning. Three measured iterations after warmup.

The audio is JFK's “We choose to go to the Moon” speech: a 20-second excerpt and the full 18-minute 15-second recording.

Speech excerpt	Warm transcription	Audio / elapsed
20 seconds	0.411 s	48.6×
18 min 15 s	11.949 s	91.7×

The long recording uses 36 chunks. Whisper-normalized WER is 42/2,220 (1.89%), matching the independent FP32 reference after normalization. A separate trace recorded 756 ANE predictions and zero GPU intervals in the measured process. That establishes placement, not ALU utilization. English JFK is the initial conversion validation; other supported languages have not yet been qualified. No word timing is supplied by these graphs.

Source .aimodel files specialize on the target device. First preparation can take several minutes. Architecture-specific AoT artifacts are omitted because no significant first-load improvement has been demonstrated. Debug source locations were removed, retaining graph signatures and operation counts. SHA256.json contains checksums of all distributed files except itself.

Device specialization and loading

The complete published bundle was measured in one preparation pass on the same Mac, after invalidating all seven assets through the public Core AI cache API:

Preparation	Time
Complete bundle, public specialization cache invalidated	3 min 0.5 s
Complete bundle, subsequent cached process	0.110 s

Preparation includes model loading and any compilation requested by Core AI, excluding warmup, transcription and the cache-invalidation call itself. The first run prepared the subsampler, all four encoder stages, both decoder assets and every supported static entry point. The cached measurement uses a new process with the same executable and assets. Both produce the same transcript as the previously qualified bundle.

The underlying ANE cache state is not fully observable. These measurements are observed on this Mac, not guaranteed fresh-install times for every device; OS/device/application changes or cache eviction can require specialization again.