In this repository, two different ASR Model will be compared in terms of converging speed, wer and cer, one is the DeepSpeech 2 and the second one is the RNN-T based Deep Speech 2.

git clone https://github.com/MorrisXu-Driving/Improving_DeepSpeech_2_by_RNN_Transducer.git
Then two .py files are under your project directory. The Deep Speech 2.py is an implementation of Deep Speech 2, an ASR Model based on DL, lauched in 2015. The DeepSpeech 2 with LSTM Transducer.py is an implementation of improved Deep Speech 2 with RNN-Transudcer in Pytorch.

This library is developed by @HawkAaron. It is a RNNT Loss function with CUDA which accelerates the training process.

cd Improving_DeepSpeech_2_by_RNN_Transducer
git clone https://github.com/HawkAaron/warp-transducer.git
cd warp-transducer

The following detailed installation can be found at https://github.com/HawkAaron/warp-transducer/

Before running the DeepSpeech 2 with LSTM Transducer.py, please make sure that comet_ml,torch,torchaudio are installed. No version restriction. If not, please run:
pip install torchaudio torch comet-ml

I ran this script on Google Cloud GPU VM with the following detailed configurations. Please compare the information below with your server accordingly before running.

The diagram shows the architecture of DeepSpeech 2. It consists of 3 ResCNN Layer and 5 Bidirectional GRU Layer and a Connectionist Temporal Classification(CTC) Decoder.

The CTC Decoder assumes that 'every output is conditionally independent of the other outputs given the input' which is not really true in sequence problems such as ASR and NLP. This results that the Char Error Rate(CER) for CTC based ASR System may be guaranteed, though, the Word Error Rate(WER) can not be guaranteed.

Therefore, in order to make past information available, RNN-Transducer is introduced.

As the graph shows that, instead of merely focus on acoustic features/spectrogram x₁-x_T, the RNN-T also regard y_u-1 as its input. Moreover, RNN-T has one more Prediction Network, namely, Decoder compared with CTC, which learn the relationship between y_u and y_u-1

This diagram shows the model architecture in this repository. Please read the code with this graph to have better understanding. The RNN-Transducer Structure consider not only the input acoustic features but also the output labels from t-1.

• Improving RNN Transducer Modeling for End-to-end Speech Recognition
• A Comparison of Sequence-to-Sequence Models for Speech Recognition
• Warp-transducer With CUDA by @HawkAaron
• Deep Speech 2: End-to-End Speech Recognition in English and Mandarin