phequals/cohere-transcribe-coreml-fp16

Cohere Transcribe — CoreML FP16

Corrected CoreML FP16 release of CohereLabs/cohere-transcribe-03-2026 for Apple Silicon inference.

This release uses:

• a length-aware encoder with inputs mel and length
• decoder prefill and decode models with explicit encoder_mask inputs
• verified token IDs: pad=2, eos=3, bos=4

The March 28 encoder fix preserves the real mel length at inference time, which avoids the padding-related repetition and hallucination issues seen with the older static encoder export.

Contents

Artifact	Precision	Notes
cohere_encoder.mlpackage	FP16	Length-aware encoder (mel, length)
cohere_decoder_prefill.mlpackage	FP16	Prompt prefill with encoder_mask
cohere_decoder_decode.mlpackage	FP16	Single-token decode with encoder_mask

Inputs / Outputs

Encoder

Input	Shape	Type
mel	[1, 128, 3500]	float32
length	[1]	int32

Output	Shape	Type
encoder_hidden	[1, 438, 1024]	float16

Decoder Prefill

Input	Shape	Type
encoder_hidden	[1, 438, 1024]	float32
input_ids	[1, 10]	int32
encoder_mask	[1, 438]	float32

Decoder Decode

Input	Shape	Type
input_ids	[1, 1]	int32
cache_update_mask	[1, 512]	float32
cache_valid_mask	[1, 512]	float32
encoder_mask	[1, 438]	float32

Prompt

English transcription with punctuation uses the token IDs:

[13764, 7, 4, 16, 62, 62, 5, 9, 11, 13]

Relevant tokenizer IDs:

• pad_token_id = 2
• eos_token_id = 3
• bos_token_id = 4

Validation

The corrected encoder was validated in Apple CoreML runtime with the same padded mel input and different length values:

• same-length repeat max diff: 0.0
• full-vs-short length max diff: 2.988

License

Apache 2.0 (same as the base model)