Yifan Yang
Yifan Yang
Yeah, I have RNNLM trained on GigaSpeech but not in icefall style. https://huggingface.co/yfyeung/icefall-asr-gigaspeech-rnn_lm-2023-10-08
@AmirHussein96 I note that you modified `k2.rnnt_loss_pruned` in k2. Would you mind sharing your branch?
**1 1** 1 0 **1 1 1** 1 **1** 1 0 0 limit_lens, which is the maximum reduced frames for each utt, is [2, 1, 3] instead of [1, 1,...
> For example, if x_lens = [10, 12], y_lens = [1, 2], then T = 12, and limit_lens = T - y_lens = [11, 10] , where limit_lens[0] = 11...
> Hi, have you got any results with phone based models? I previously tried with librispeech and the result was worse than BPE. For pruned transducer I only got 4-5...
> Maybe sometime later. Not recently.
@drawfish Thanks for your suggestion. This model is in the LibriSpeech, whose test set does not have entirely silent sentences. IMO, you should modify the code of the export model.
> I'm moving the conversation here since the previous [PR](https://github.com/k2-fsa/icefall/pull/1500) was closed. > > I ran @yfyeung 's training command using the merged k2ssl codes, with the batch-size and world-size...
> ```shell > --max-duration 300 \ > --accum-grad 4 \ > ``` The current gradient accumulation mechanism simulates multi-GPU setup. You can simulate my setup using 4 GPUs with `acc_grad`...