VisualEars FastConformer Persian ASR ONNX FP16
Real FP16 fixed-frame ONNX CTC core for browser/WebGPU experiments.
This artifact expects precomputed 80-bin log-mel features with shape [1, 80, 2005] and dtype float16. It outputs logits as float16 with shape [1, 251, 1025] plus encoded_lengths as int64.
Files:
fastconformer_ctc_fixed2005_fp16_full_io.onnxfastconformer_ctc_fixed2005_fp16_full_io.onnx.data
Validation against the fp32 ONNX source on VisualEars269:
{
"n": 269,
"exact_norm_match_rate": 1.0,
"collapsed_match_rate": 1.0,
"sequence_argmax_equal_rate": 0.9888475836431226,
"mean_frame_argmax_match_rate": 0.9999555680623231,
"wer_vs_fp": 0.0,
"cer_vs_fp": 0.0
}
Checksums:
735fad7c997ad57c90918715655718099497133b4ab2dc720c5dc291f2a70ce3 fastconformer_ctc_fixed2005_fp16_full_io.onnx
462f2fb5408c7deb8ee969ea999d50208a5fc965c5fb1eb9855695fb687a0082 fastconformer_ctc_fixed2005_fp16_full_io.onnx.data
Model tree for Reza2kn/visualears-fastconformer-fa-full-ab-onnx-fp16
Base model
nvidia/stt_fa_fastconformer_hybrid_large