VisualEars FastConformer Persian ASR CoreML W4 Quality Variant

Derivative export of Reza2kn/visualears-fastconformer-fa-full-ab.

Artifact

  • Format: CoreML selective 4-bit k-means quality-first export
  • Quantization/conversion: CoreML palettize_weights, k-means, nbits=4, weight_threshold=1100000
  • Runtime validation: CoreML CPU
  • Validation result: 99.65% CTC argmax parity
  • Size: 217 MB, 99.1% of CoreML FP16 source

Validation

Runtime parity was checked against PyTorch CTC logits on 16 calibration clips padded to 2005 mel frames. The metric is CTC argmax token agreement versus the PyTorch reference logits, not end-to-end WER.

Usage Boundary

These are fixed-frame acoustic CTC-core exports. They take precomputed log-mel features as processed_signal; they are not full raw-audio-to-text pipelines by themselves.

Notes

Quality-first CoreML W4 variant. It passes with a wider margin but barely compresses.

Downloads last month
4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Reza2kn/visualears-fastconformer-fa-full-ab-coreml-w4-quality