CoolFace
Modelpublic

Reza2kn/visualears-fastconformer-fa-depoisoned-phaseB-onnx-fp16

sourceHugging Facecc-by-nc-4.0updated 3mo agoView on Hugging Face
0likes
Model Card

VisualEars PhaseB FastConformer ONNX FP16

Browser-oriented FP16 ONNX CTC-core export for Reza2kn/visualears-fastconformer-fa-depoisoned-phaseB, using fixed 2005 log-mel frames and processed_signal_length.

Files:

  • —fastconformer_phaseB_ctc_fixed2005_len_fp16_full_io.onnx
  • —fastconformer_phaseB_ctc_fixed2005_len_fp16_full_io.onnx.data
  • —fastconformer_phaseB_ctc_fixed2005_len_fp16_full_io_embedded.onnx
  • —tokens.json
  • —preprocessor.json
  • —mel_filters_slaney_80x257.json

Validation against PhaseB FP32 ONNX on the VisualEars269 browser-feature gate:

json
{
  "n": 269,
  "baseline_model": "artifacts/phaseB-ctc-fp16-len/fastconformer_fa69m_ctc_fixed2005_len_fp32_full_io.onnx",
  "candidate_model": "artifacts/phaseB-ctc-fp16-len/fastconformer_phaseB_ctc_fixed2005_len_fp16_full_io_castfixed.onnx",
  "baseline_size_bytes": 460523172,
  "candidate_size_bytes": 231640784,
  "size_ratio_vs_baseline": 0.5029948503872461,
  "exact_norm_match_rate": 0.9925650557620818,
  "collapsed_match_rate": 0.9925650557620818,
  "sequence_argmax_equal_rate": 0.9851301115241635,
  "mean_frame_argmax_match_rate": 0.9996373197932723,
  "char_similarity_to_baseline": 0.9996485679142506,
  "mean_abs": 0.006820719505636,
  "max_abs": 1.872690200805664
}