| { |
| "model_type": "fela-asr-ctc", |
| "architectures": [ |
| "FelaAsrModel" |
| ], |
| "auto_map": { |
| "AutoConfig": "configuration_asr.FelaAsrConfig", |
| "AutoModel": "modeling_asr.FelaAsrModel" |
| }, |
| "architecture": "FELA", |
| "edition": "streaming 64M", |
| "n_layer": 16, |
| "n_embd": 512, |
| "n_head": 8, |
| "fno_modes": 256, |
| "gla_chunk": 64, |
| "ffn_hidden": 2048, |
| "layer_pattern": "FNO", |
| "dropout": 0.0, |
| "bias": false, |
| "gla_delta": false, |
| "n_mels": 80, |
| "sample_rate": 16000, |
| "n_fft": 400, |
| "win_length": 400, |
| "hop_length": 160, |
| "subsample_factor": 4, |
| "vocab": 257, |
| "bpe_units": 256, |
| "ctc_blank_id": 0, |
| "decode": "greedy ctc", |
| "language_model": "none", |
| "tokenizer": "bpe256.model (SentencePiece, 256 units + 1 CTC blank)", |
| "note": "The 64.4M streaming model (test clean WER 16.77 at 460h). The 206M byte CTC model (test clean 14.3 / test other 27.4 at 960h) shares the architecture with a larger width/depth." |
| } |
|
|
