Speech Recognition for Uyghur using Speech transformer

Training:

this model using CTC loss and Cross Entropy loss for training.

Download pretrained model and dataset.

unzip results.7z and thuyg20_data.7z to the same folder where python source files located. then run:

python train.py

Recognition:

for recognition download only pretrained model. then run:

python .\tonu.py .\test6.wav

result will be:

        Model loaded: results/UFormer_last.pth
            Best CER: 4.16%
             Trained: 276 epochs
The model has 36,418,306 trainable parameters
 Feature  has 25,869,058 trainable parameters
  Encoder has 4,205,568 trainable parameters
  Decoder has 6,343,680 trainable parameters

======================
Recognizing file .\test6.wav
test6.wav -> u qizlarning resimi chiqip qalsa bilekchila sinchilap qaraytti

This project using

A free Uyghur speech database Released by CSLT@Tsinghua University & Xinjiang University

Reference

https://github.com/gentaiscool/end2end-asr-pytorch