haloop on LibriSpeech

This progress report shows training runs of haloop's LSTM acoustic and language models.

Created on April 16|Last edited on April 16

Comment

﻿haloop is my speech agent toolkit: pip install haloop
The toolkit code is available on GitHub: https://github.com/proger/haloop﻿
This toolkit is designed to accompany teaching materials for Waves and Symbols tutorial: https://www.overleaf.com/read/dhhnthbcbppq﻿
Hac LSTM acoustic model is trained using CTC to predict a CMUdict phoneme sequence given MFCC inputs on train-clean-360 and train-other-500. train-clean-100 is generally reserved for finetuning experiments. dev-clean is used to perform validation.
﻿
valid/loss
valid/loss
Showing first 10 runs
5001k1.5kStep12345
valid/ler
valid/ler
Showing first 10 runs
5001k1.5kStep0.20.40.60.811.21.4
Run set622
﻿
Hal LSTM language model is trained to predict the next byte (256 outputs) using corpus librispeech-lm-norm.txt from https://www.openslr.org/11/  
﻿
Run set2
﻿
﻿

Add a comment