Hello. Im trying to train Russian model on Deepspeech 0.6.1 and face with really high loss.
I get corpus from here, its about 72 hours of speech. Then i convert mp3 to wav with import_cv2.py, in result i have:
train.csv - 9743 wavs
dev.csv - 5714 wavs
test.csv - 5892 wavs
Then i start training by this arguments:
python -u DeepSpeech.py
–train_files /media/djo/DS/work/rumodel/clips/train.csv
–dev_files /media/djo/DS/work/rumodel/clips/dev.csv
–test_files /media/djo/DS/work/rumodel/clips/test.csv
–n_hidden 375
–epochs 10
–learning_rate 0.00095
–export_dir data/export
–checkpoint_dir data/cp_ru_full
In result i have:
I Early stop triggered as (for last 4 steps) validation loss: 131.891235 with standard deviation: 0.253853 and mean: 127.696702
Why so huge loss?