icefall

icefall copied to clipboard

Zipformer Training Issues:Gradient too small and cuda out of memory issue

9

Hello All, we are training a zip former model for about 3400 hours of Tamil data. We were facing this issue: RuntimeError: grad_scale is too small, exiting: 1.4901161193847656e-08 We have...

bsshruthi22

k2SSL: A Faster and Better Framework for Self-Supervised Speech Representation Learning

1

Paper link: https://arxiv.org/abs/2411.17100v1

yfyeung

Early Stopping of Token Generation in Streaming Model Training

27

Hi Next-gen Kaldi team, Thank you once again for your continuous support and patience with our Japanese ASR recipe and model developments. We're currently training the streaming model based on...

Triplecq

wgfi110

https://github.com/k2-fsa/icefall/blob/3b257dd5ae79bff99470ec1cbbeaa8fae84f956a/egs/wenetspeech/KWS/prepare.sh#L66 As shown in the figure below, the path of `open-commands->scripts` is incorrect, this MR just fix open-commands path in kws. 

yulianjie

JinZr

devision by zero

2

Airgods

Hi, I'm confused about the pad_length in set_features. Since the frames that are less than chunk_size\*2+7+2\*3 have already been padded in streaming_decode.py, why are we still adding 7+2\*3 when initially...

1215thebqtic

No duration or confidence info for alignments

2

I've tried generating alignments for a `pruned_transducer_stateless7` model using https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py. Looking at the output cuts I can only find the start times for tokens/words. Is there a way to get...

AhmedSalah98

fix the CTC zipformer2 training

10

- too many supervision tokens - change filtering rule to `if (T - 2) < len(tokens): return False` - this prevents inf. from appearing in the CTC loss value (empirically...

KarelVesely84