Wave2Letter (paper) with a waveform input.
The model was trained on LibriSpeech-960. In training, BatchNorm and Dropout were used, which can be fused into the weights to make them compatible with the Wave2Letter from torchaudio.models.
Pretrained weights
for model.Wav2Letter (link)
for torchaudio.models.Wav2Letter (link)
Greedy decoding
| dataset | CER | WER |
|---|---|---|
| dev-clean | 0.111 | 0.331 |
| test-clean | 0.105 | 0.318 |
Example
from torchaudio.models import Wav2Letter
model = Wav2Letter(num_classes=len(labels)).cuda()
model.load_state_dict(torch.load('./pretrained/states_fused.pth'))Some filter kernels from the first Conv1d layer
