icefall

icefall copied to clipboard

fine-tune problem

14

When I tried fine-tuning, I found that the WER value after fine-tuning was over 100, which should be a problem with the fine-tuning. Below is a simple fine-tuning log document...

PeaceAndJoyAaron

Non-streaming Conformer model with pruned_rnnt_loss always emits the first non-blank characters on the very first frames.

1

I trained two offline reworked conformer models on my own Chinese data using pruned_rnnt_loss and standard rnnt loss (warp-rnnt==0.7.0) respectively following pruned_transducer_stateless5. However, I still experience the issue where the...

guoyifan97

This PR implements https://arxiv.org/pdf/2109.00627

pkufool

fast_beam_search_nbest gives very high WER compared to fast_beam_search and greedy.

2

Hi, I am getting around 3% wer in fast-beam-search and greedy-search. However, I am getting 70% WER when I use fast-beam-search-ngram. My decode configuration looks as below. I am using...

chirag-augnito

This PR uses pinyin as modeling units for Chinese.

pkufool

2024-06-20 09:55:25,806 INFO [train.py:1256] (4/8) Saving batch to exp/pts7_aatmd4k_ctc_bs/batch-c33f4584-b23b-c1d8-493c-d01609de8895.pt 2024-06-20 09:55:25,868 INFO [train.py:1262] (4/8) features shape: torch.Size([157, 477, 80]) 2024-06-20 09:55:25,870 INFO [train.py:1267] (4/8) num tokens: 2872 Traceback (most recent...

kellkwang

Hi, when I training a zipformer model using the official configuration, I find the right chunk size is about 130ms. It maybe too long for streaming ASR, Is there any...

csf123123

To prepare phone based lang, I see generate_unique_lexicon.py is used in almost every Chinese ASR eg(e.g. aishell-*), but it's not in English ASR(e.g. gigaspeech, librispeech), what's the reason? I want...

yzchen563

kaldifeat installation error

2

Hello guys, I have following kaldifeat installation error. One week ago everything was OK with this installation. Any help is more than appreciated. (base) root@gpu-alex-0-0:/workspace/inputs/alexg/asr/src/models/k2_2024/icefall/egs/tedlium3/ASR/sh# pip install kaldifeat==1.25.1.dev20231022+cuda11.6.torch1.13.0 -f https://csukuangfj.github.io/kaldifeat/cuda.html...

AlexandderGorodetski

Questions about modifying prepare.sh for training ASR model on custom data

2

Hi, opening a new issue since the old one has been [closed](https://github.com/k2-fsa/icefall/issues/1627). Currently, we are writing our own `prepare.sh` to train an ASR model based on our own Chinese audio...

daocunyang