Upload 115m_preprocessor.json with huggingface_hub
Browse files- 115m_preprocessor.json +69 -0
115m_preprocessor.json
ADDED
|
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema": "visualears-preprocessor-v2-nemo-fbank",
|
| 3 |
+
"description": "NeMo AudioToMelSpectrogramPreprocessor-compatible browser feature pipeline for VisualEars PhaseB. This model uses normalize=per_feature, no preemphasis, centered STFT, Slaney mel filters, and requires processed_signal_length.",
|
| 4 |
+
"sample_rate": 16000,
|
| 5 |
+
"n_fft": 512,
|
| 6 |
+
"win_length": 400,
|
| 7 |
+
"hop_length": 160,
|
| 8 |
+
"n_mels": 80,
|
| 9 |
+
"window": "hann_periodic_false",
|
| 10 |
+
"window_formula": "0.5 - 0.5*cos(2*pi*i/(win_length-1))",
|
| 11 |
+
"center": true,
|
| 12 |
+
"center_pad": 256,
|
| 13 |
+
"pad_mode": "reflect",
|
| 14 |
+
"preemphasis": null,
|
| 15 |
+
"mel_scale": "slaney/librosa.filters.mel(htk=False,norm='slaney')",
|
| 16 |
+
"mel_filters_file": "mel_filters_slaney_80x257.json",
|
| 17 |
+
"spectrum": "magnitude_power_2_no_fft_normalization",
|
| 18 |
+
"log": "natural",
|
| 19 |
+
"log_zero_guard_type": "add",
|
| 20 |
+
"log_zero_guard_value": 5.960464477539063e-08,
|
| 21 |
+
"normalize": "per_mel_bin_mean_variance_over_valid_frames",
|
| 22 |
+
"fixed_frames": 2005,
|
| 23 |
+
"pad_value": 0.0,
|
| 24 |
+
"frame_count_formula": "max(1, min(fixed_frames, floor(num_samples / hop_length) + 1))",
|
| 25 |
+
"inputs": [
|
| 26 |
+
{
|
| 27 |
+
"name": "processed_signal",
|
| 28 |
+
"dtype": "float32",
|
| 29 |
+
"shape": [
|
| 30 |
+
"batch",
|
| 31 |
+
80,
|
| 32 |
+
2005
|
| 33 |
+
]
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"name": "processed_signal_length",
|
| 37 |
+
"dtype": "int64",
|
| 38 |
+
"shape": [
|
| 39 |
+
"batch"
|
| 40 |
+
],
|
| 41 |
+
"description": "valid log-mel frame count before zero padding"
|
| 42 |
+
}
|
| 43 |
+
],
|
| 44 |
+
"outputs": [
|
| 45 |
+
{
|
| 46 |
+
"name": "logits",
|
| 47 |
+
"dtype": "float32",
|
| 48 |
+
"shape": [
|
| 49 |
+
"batch",
|
| 50 |
+
251,
|
| 51 |
+
1025
|
| 52 |
+
]
|
| 53 |
+
},
|
| 54 |
+
{
|
| 55 |
+
"name": "encoded_lengths",
|
| 56 |
+
"dtype": "int64",
|
| 57 |
+
"shape": [
|
| 58 |
+
"batch"
|
| 59 |
+
]
|
| 60 |
+
}
|
| 61 |
+
],
|
| 62 |
+
"output_stride": 8,
|
| 63 |
+
"usable_steps_formula": "min(encoded_lengths[0], logits_steps)",
|
| 64 |
+
"ms_per_output_step": 80,
|
| 65 |
+
"blank_id": 1024,
|
| 66 |
+
"ctc_decode": "greedy argmax; drop blank, repeats, and SentencePiece special tokens like <pad>/<unk>/<fa-IR>; join pieces; '▁' -> space",
|
| 67 |
+
"normalize_source": "NeMo AudioToMelSpectrogramPreprocessor normalize=per_feature",
|
| 68 |
+
"normalize_eps": 1e-05
|
| 69 |
+
}
|