Welm

WeLM

WeLM depens on:

1. WeTextProcessing: Text Normalization
2. jieba: Chinese Word Segmentation
3. SentencePiece : English subword units Byte-Pair-Encoding (BPE)
4. SRILM: N-Gram Language Model Training
5. Some tools of Kaldi: TLG Graph Building

Usage

1. Prepare the acoustic model's unit file units.txt.

$ wget https://path/to/aishell2/20210618_u2pp_conformer_libtorch.tar.gz
$ tar -xf 20210618_u2pp_conformer_libtorch.tar.gz

2. Prepare transcript file trans.txt.

$ head -n1 trans.txt
BAC009S0002W0122 而对楼市成交抑制作用最大的限购

3. Build the FST.

$ bash run.sh --unit_file units.txt --text trans.txt
$ ls exp/lang