dynamic-asr / real /config.json
UncleanCode's picture
Dynamic ASR v0.1.0: real (default) + codeswitch-experimental
e4c97d2 verified
Raw History Blame Contribute Delete
1.97 kB
{
"base_model": "omniASR_CTC_1B_v2 (facebook omnilingual-asr, Apache-2.0)",
"model_card": "omniASR_CTC_1B_v2",
"architecture": "wav2vec2 CTC 1B (unchanged)",
"vocab_size": 10288,
"tokenizer": "omniASR_tokenizer_written_v2",
"checkpoint_format": "torch.save dict {model: state_dict (bf16), step, dev, card}",
"sample_rate": 16000,
"decoding": "greedy CTC, Latin-script logit mask, no language hint",
"languages": [
"en",
"pcm",
"ig",
"yo",
"ha"
],
"license": "cc-by-sa-4.0",
"training": {
"steps": 4000,
"batch_audio_sec": 160,
"max_clip_sec": 18,
"lr": 2e-05,
"warmup": 300,
"schedule": "linear decay",
"optimizer": "AdamW b(0.9,0.98) wd 0.01",
"precision": "bf16 autocast",
"frozen": "CNN feature encoder",
"seed": 1,
"real_aug_p": 0.15,
"synthetic_codeswitch_fraction": 0.0,
"data": "licensed real Nigerian speech only"
},
"name": "dynamic-asr/real",
"status": "default",
"file": "model.pt",
"scores": {
"duplex_eng": {
"n": 29,
"wer": 0.0953,
"cer": 0.0252,
"cer_nodiac": 0.0252
},
"duplex_pcm": {
"n": 40,
"wer": 0.1927,
"cer": 0.0738,
"cer_nodiac": 0.0736
},
"duplex_ibo": {
"n": 40,
"wer": 0.3153,
"cer": 0.0929,
"cer_nodiac": 0.0761
},
"duplex_yor": {
"n": 40,
"wer": 0.3547,
"cer": 0.1213,
"cer_nodiac": 0.0841
},
"duplex_hau": {
"n": 40,
"wer": 0.2533,
"cer": 0.0823,
"cer_nodiac": 0.0823
},
"duplex_cs": {
"n": 40,
"wer": 0.2385,
"cer": 0.0807,
"cer_nodiac": 0.0646
},
"realcs_all": {
"n": 94,
"wer": 0.4357,
"cer": 0.1482,
"cer_nodiac": 0.137
},
"realcs_natural_conversational": {
"n": 36,
"wer": 0.6227,
"cer": 0.2481,
"cer_nodiac": 0.2335
},
"realcs_insertional": {
"n": 58,
"wer": 0.3648,
"cer": 0.117,
"cer_nodiac": 0.107
},
"realcs_eng_word_recall": 0.3186,
"synthtest": {
"n": 109,
"wer": 0.2932,
"cer": 0.1052,
"cer_nodiac": 0.0914
}
}
}