visualears-fastconformer-fa32m-streaming-bpe1024-onnx-fp16

VisualEars FastConformer FA32M Streaming BPE1024 — ONNX FP16 WebGPU

Derivative export of Reza2kn/visualears-fastconformer-fa32m-streaming-bpe1024.

This is the length-aware FA32M CTC acoustic-core export. It fixes the blank-transcript failure from one-input fixed2005 exports by requiring the real valid feature length.

Artifact

• Format: ONNX FP16 WebGPU fixed-frame acoustic CTC-core export
• Runtime validation: ONNX Runtime CPU
• Validation result: 267/269 exact transcript matches (99.26%), argmax step agreement 99.9824%
• ONNX non-empty transcript rate: 98.88% on the 269-clip difficult short/noisy set
• Size: 116.4 MiB ONNX artifacts

Runtime contract

• input: processed_signal float16 [batch, 80, 2005]
• input: processed_signal_length int64 [batch] — valid log-mel frame count before zero padding
• output: logits float16 [batch, 252, 1025]
• output: encoded_lengths int64 [batch]
• opset: 18, IR version: 10

Files

Validation

Source: PyTorch NeMo preprocessor + encoder + auxiliary CTC decoder in fp32.
Candidate: this ONNX fp16 artifact in ONNX Runtime CPUExecutionProvider.

Validation set: all 269 clips from Reza2kn/visualears-benchmark-269-gold.

{
  "created_at_utc": "2026-06-14T23:44:55Z",
  "clips": 269,
  "exact_transcript_matches": 267,
  "exact_transcript_parity": 0.9925650557620818,
  "source_nonempty": 266,
  "onnx_nonempty": 266,
  "onnx_nonempty_rate": 0.9888475836431226,
  "mean_char_similarity": 0.9996047132394416,
  "min_char_similarity": 0.9259259259259259,
  "argmax_step_agreement": 0.9998240520805841,
  "wall_sec": 45.74090790748596,
  "source": "PyTorch NeMo preprocessor + encoder + auxiliary CTC fp32",
  "candidate": "ONNX fp16 encoder + auxiliary CTC with processed_signal_length"
}

Browser / runtime feature contract

Use the sidecars in this repo:

• tokens.json — CTC vocabulary and special token ids
• preprocessor.json — exact FA32M NeMo fbank contract
• mel_filters_slaney_80x257.json — Slaney/librosa mel filters

Feature summary:

• preemphasis: y[0]=x[0], y[t]=x[t]-0.97*x[t-1]
• STFT: n_fft=512, win_length=400, hop=160, center=True, reflect pad by 256 samples, Hann periodic=false centered inside the FFT frame
• spectrum: unnormalized FFT magnitude squared
• mel: Slaney/librosa filterbank from the sidecar
• log: ln(mel_energy + 2^-24)
• no per-bin normalization (normalize=NA)
• zero-pad features to 2005 frames and pass the true processed_signal_length
• greedy CTC: use encoded_lengths[0], drop blank, repeats, and SentencePiece special tokens like <pad>, <unk>, <fa-IR>

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.

Provenance

• Source repo: Reza2kn/visualears-fastconformer-fa32m-streaming-bpe1024
• Source file: fa32m_streaming_bpe1024_final.nemo
• Source SHA256: 034fb2afa19da13db8a120970a7f8d3e696987014cc62684ce50a1382d332448
• Architecture: EncDecHybridRNNTCTCBPEModel encoder + auxiliary CTC decoder with processed_signal_length input