This is a branch of Shijie Wu's neural-transducer
library
- Environment (conda):
environment.yml
- Pre-commit check:
pre-commit run --all-files
- Compile:
make
If using pretrained model checkpoints, create a checkpoints
directory and unzip binaries, e.g.,
mkdir checkpoints
unzip models.zip -r checkpoints/
After compiling, you can run an example of phoneme to word (graphemes) using the following command
bash p2word_decode.sh
which uses phonemes in cmudict/input_example.tsv
as an example. New models can be trained using the trm-p2word.sh script in examples/transformer
where training data can be pulled/processed using the script in the cmudict
directory. Directory names in the p2word_decode script may need to be changed accordingly.