Codecov Report

Merging #4944 (1835a92) into master (9c7bde4) will increase coverage by 0.01%.
The diff coverage is 86.88%.

@@            Coverage Diff             @@
##           master    #4944      +/-   ##
==========================================
+ Coverage   76.63%   76.65%   +0.01%     
==========================================
  Files         604      604              
  Lines       53934    53992      +58     
==========================================
+ Hits        41334    41385      +51     
- Misses      12600    12607       +7     

Flags with carried forward coverage won't be shown. Click here to find out more.

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

I changed my mind.

I implemented --filtered_train_key_text and --filtered_valid_key_text for espnet2/bin/*_train.py.

With giving a text file containing IDs to be filtered,
The samples specified by this option are excluded from the training.

e.g.

• wav.scp:

IDa a.wav
IDb b.wav
IDc b.wav

• filtered_key.txt

IDb

In this case, IDb is excluded, and IDa and IDc remain for the training.

I also changed asr.sh to create filtered_key.txt at stage 4 instead of creating a new dataset.

I changed my mind again.

Filtering short/long utterances by the option of the python tool is better way as a viewpoint for a smart recipe, but it could make some overhead for start at the startup.

Creating another dataset is a dirty way, but actually efficient for training speed.

I'll think about it.