voxor-ai/cohere-transcribe-arabic-coreml

Cohere Transcribe Arabic audio encoder, Core ML

The audio encoder of CohereLabs/cohere-transcribe-arabic-07-2026 as compiled Core ML models (AudioEncoder/chunk00.mlmodelc, chunk01.mlmodelc, chunk02.mlmodelc, run in that order; int8 weights, fp16 compute, iOS 17 / macOS 14 or later).

• chunk00 inputs: mel [1, 128, 3000] float32 (30 s of log-mel frames), length [1] int32 (valid mel frames); output hidden_out [1, 375, 1280] float32
• chunk01 and chunk02 inputs: hidden [1, 375, 1280] float32 (the previous part's hidden_out), key_mask [1, 1, 1, 375] float32 (0 for a valid row, -10000 beyond), pad_mask [1, 375, 1] float32 (1 for a valid row, 0 beyond); outputs hidden_out [1, 375, 1280] and, for chunk02, audio_features [1, 375, 1024] float32
• Valid rows: n = floor((floor((floor((length - 1) / 2) + 1 - 1) / 2) + 1 - 1) / 2) + 1
• Files over 1 MB are zstd-compressed (<file>.zst); decompress before use.

Accuracy: output cosine similarity to the unconverted encoder 0.9996 mean, 0.9993 minimum. On FLEURS Arabic (60 clips) transcripts made with this encoder score 4.02 % WER, the same as the unconverted encoder in the same pipeline.

Apache License 2.0; see LICENSE and NOTICE.