Lightweight-Transducer

Official implementation of the Interspeech 2024 paper "Lightweight Transducer Based on Frame Level Criterion".

Install

• Python3
• pip install -r requirements.txt
• (Optional) Refer to https://kheafield.com/code/kenlm/ for installing KenLM.

Usage

Data preparation

Download AISHELL-1 and extract it to the directory data/aishell.

python data/aishell/get_csv.py
python data/aishell/get_vocab.py

Training

python train.py

Parameter averaging

python avg_model.py

Test

python test.py

Speed

Training one epoch takes about five minutes with a single GPU RTX 4090 and CPU i9-13900K.

Results

Testset	Sub	Del	Ins	CER
dev	3.79	0.10	0.07	3.96
test	4.10	0.16	0.05	4.31

Pre-trained model

Rescore

Data preparation

Download resource_aishell and extract it to the directory data/aishell.

python data/aishell/get_text.py

Training

../kenlm/build/bin/lmplz -o 3 --text data/aishell/aishell_train.txt --arpa data/aishell/aishell_train.arpa -S 10% --interpolate_unigrams 0

Pre-trained language model

Rescore

python rescore.py

Results

Testset	Sub	Del	Ins	CER
dev	3.61	0.10	0.07	3.78
test	3.82	0.16	0.04	4.03