noahchalifour / baidu-deepspeech2

A Tensorflow implementation of Baidu's Deep Speech 2 paper

Baidu's Deep Speech 2 (Tensorflow)

(This is a work in progress)

This is a python implementation of Baidu's Deep Speech 2 paper https://arxiv.org/pdf/1512.02595.pdf using tensorflow

TODO:

• Fix GPU memory
• Add batch normalization to RNN
• Implement row convolution layer
• Add other dataset support
• Create pretrained models

Preprocessing

To preprocess your data you must first download the one of the datasets above and extract them to a folder. Then run the following script to preprocess the data (This might take a while depending on the amount of data you have)

python preprocess.py --data-dir=<your data directory> --dataset=<dataset name>

Training

Now that you have preprocessed your data, you can train a model. To do this, you can edit the settings in the config.py file if you want. Then run the following command to train the model:

python train.py

Testing your model

Now that you have trained a model, you can go ahead and start using it. We have created two scripts that can help you do this infer.py and streaming_infer.py. The infer.py script, transcribes a audio file that you give it

python infer.py -f <your audio file name>

The streaming_infer.py script uses PyAudio to record audio from your computer's microphone and transcribes it in real-time. To run it simply:

python streaming_infer.py

baidu-deepspeech2's Issues

Why log_linear_specgram?

The paper does not make reference to a log_linear spectrogram.

How to predict unlabeling test data?

Hi. I'm Studying Speech Recognition. I have some question.
This model takes 4 inputs, below code.

self.model = tf.keras.Model(inputs=[input_data, labels, input_length, label_length], outputs=[loss_out])

If this model predicts test case that is unlabeled, i can't give 'labels', 'label_length' as inputs.
How can i do?

Using pretrained model without downloading the data

Is it possible to directly use the pre-trained model without downloading the whole 20-30 GBs of data?

What accuracy are you able to attain with this?

And how does it compare to the reference from the paper?

How to evaluate and test this? Infer.py and stream_infer.py files not available

Hi, I am studying ASR and its implementation. Here I have done upto training part, now want to evaluate and test the model. How can I do this, as there is no evaluate and testing file is available..
Thanks.

only one epoch is ok

Contact info

Hi Noah,

I am an undergraduate student and I am using your project as a guideline for creating a similar speech recognition system. I have a few questions, however, I had not been able to contact you. Please contact me on my email: [email protected] or on facebook - Koko Tonev. I found your profile on Facebook and messaged you but I got no response and I have some problems with my system that I don't understand... Many thanks in advance!

Kind regards,
Koko Tonev