🚧 In Work 🚧

Speech to text

A project of Speech Recognition in python using Tensorflow and keras.

Model Architecture and Training method

I train the model using a network of recurrent neurons predicting a linear output. my dataset consists of audio associated with a sentence. I cut each sentence into phonemes to which I come to associate a sound.

Model Architecture

The architecture is the following:

• One convolution of 8 filters (9*9) [Elu activation]
• Max pooling pool_size=[2,2]
• Lstm of 128 filters
• Flatten layer
• Dropout: 0.4
• Fully conected: 256 [Elu]
• Dropout: 0.2
• Fully conected: len(vocab)