---
license: cc-by-nc-4.0
base_model: Reza2kn/visualears-fastconformer-fa-depoisoned-phaseB
base_model_relation: quantized
language:
- fa
pipeline_tag: automatic-speech-recognition
tags:
- visualears
- fastconformer
- streaming
- mlx
- fp16
- rnnt
- ctc
---

# VisualEars PhaseB Streaming MLX FP16

MLX FP16 weight bundle and RNNT runtime scaffold for `Reza2kn/visualears-fastconformer-fa-depoisoned-phaseB`.

Files:

- `phaseB_hybrid_streaming_fp16.safetensors`
- `phaseB_hybrid_streaming_fp16_mlx_manifest.json`
- `phaseb_mlx_rnnt_runtime.py`
- `mlx_rnnt_runtime_smoke.json`
- `mlx_vs_torch_rnnt_argmax20.json`

Validation: MLX RNNT predictor/joint smoke passed; RNNT joint argmax matched TorchScript on 20 random probes.

The MLX encoder port is not complete in this artifact set. The exported weights and manifest preserve the streaming cache/head metadata for the full hybrid model.