icefall
icefall copied to clipboard
k2-fsa
This is a Librispeech zipformer recipe using HAT loss from https://github.com/k2-fsa/k2/pull/1244. The recipe includes HAT training, greedy decoding, modified beam search decoding, and subtracting ILM with RNN-LM shallow fusion. So...
PyTorch 2.0 is now available which now includes two major things that may be useful for faster training: 1. TorchDynamo, which can be used by calling `torch.compile(model)` 2. FLASH attention,...
Hi, I have trained zipformer2 (without streaming) model with my dataset. Training command: **./zipformer/train.py --num-epochs 40 --start-epoch 1 --use-fp16 1 --enable-musan False --exp-dir zipformer/exp-small --causal 0 --num-encoder-layers 2,2,2,2,2,2 --feedforward-dim 512,768,768,768,768,768...
Keeping a separate vector as an nn.Parameter in the model, and adding it times k/(1+k) to the embedding before the ReLU during training and decoding. It only supports greedy_search and...
what are the steps to train the icefall asr model in aws sagemaker
Note that this PR does not include recipes for non-discrete features. The paper link: https://arxiv.org/pdf/2309.07377.pdf
I'm getting 77% wer. %WER 77.58% [21709 / 27984, 465 ins, 13810 del, 7434 sub ] how to reduce the wer and what will be the reason for getting high...
I am playing a bit with the CTC option in the zipformer 2, with the largest model from the documentation. It trained well for a first dataset but when I...
Please see https://arxiv.org/abs/2310.11230 [2310.11230.pdf](https://github.com/k2-fsa/icefall/files/13063856/2310.11230.pdf)
This is a pull request for Dialectal IWSLT-Tunisian 2022 shared task https://iwslt.org/2022/dialect ASR and ST recipes.