Project42/wav2vec2-large-lv60-align

wav2vec2-large-960h-lv60-self — alignment ONNX (fp16)

ONNX export of facebook/wav2vec2-large-960h-lv60-self for CTC forced alignment in the browser (onnxruntime-web, WebGPU EP):

• input input_values [1, samples] float32 @16 kHz (zero-mean/unit-var normalized host-side)
• output logits [1, frames, 32] — the standard wav2vec2-960h char vocab (<pad>=0 = CTC blank, |=4 word separator, A–Z + ')

fp16-converted (fp32 I/O). Export + conversion script: eval/export-large-align.py in the KaraokeVideoEditor repo.

Chosen over base-960h for lyrics alignment: the large model distinguishes repeated chorus occurrences (kiri benchmark: p90 247 ms vs 8.3 s for base).