CoolFace
Modelpublic

Reza2kn/visualears-fastconformer-fa-full-ab-onnx-w2-progressive

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
0likes
Model Card

VisualEars FastConformer FA W2 Progressive ONNX

Fixed-frame ONNX CTC acoustic core exported from the progressive 2-bit QAT materialized checkpoint:

onebit_progressive_l0_fast_no_urls_lower3_20260612_170013/fa_fastconformer_qat_w2_cotrained_materialized.nemo

This is the same browser-facing interface as the FP demo:

  • input: processed_signal, float32 [batch, 80, 2005]
  • outputs: logits [batch, 251, 1025], encoded_lengths [batch]
  • external data: fastconformer_ctc_fixed2005.onnx.data

Important: this ONNX is exported from a materialized QAT checkpoint. The selected weights are projected to W2 values, but this is not bit-packed/compressed yet.

VisualEars269 score for the source W2 materialized checkpoint:

decodeWERCER
default36.92%17.07%
forced CTC39.67%17.86%