Reza2kn/visualears-fastconformer-fa-depoisoned-phaseB-streaming-coreml-fp16

VisualEars PhaseB Streaming CoreML FP16

CoreML MLProgram FP16 streaming split for Reza2kn/visualears-fastconformer-fa-depoisoned-phaseB.

Files:

• streaming_encoder_ctc_streaming_fp16.mlpackage
• rnnt_predictor_streaming_fp16.mlpackage
• rnnt_joint_streaming_fp16.mlpackage
• coreml_fixed_streaming_manifest.json

Streaming contract:

• feature chunk: 1 x 80 x 32
• chunk/shift config: chunk_size=4, shift_size=3, left_chunks=64
• feature shift window: [17, 24]
• encoder output per chunk: 1 x 512 x 2
• CTC logits per chunk: 1 x 2 x 1025
• channel cache: 17 x 1 x 256 x 512
• time cache: 17 x 1 x 512 x 4
• heads: RNNT and CTC

Validation so far:

• All three .mlpackage files load and run via CoreML CPU execution.
• Streaming encoder+CTC CTC argmax matched TorchScript on smoke probes.
• RNNT joint argmax matched TorchScript on smoke probes.
• Limit-5 streaming CTC transcript gate matched TorchScript source exactly.
• Full 269 streaming CTC transcript gate matched TorchScript streaming source on 265/269 normalized transcripts: 98.513%.
• Mean frame argmax match rate on the 269 gate: 99.946%.

Implementation note: the CoreML encoder uses a fixed streaming-contract shim for CoreMLTools conversion. It keeps the cache-bearing streaming interface but replaces NeMo runtime shape-to-Python-int slices with constants for this exact contract.