This repository contains a Tensorflow implementation of Baidu's Deepspeech architecture, a neural speech-to-text model. This repository was inspired by the Mozilla's implementation.
The model architecture is a sequence-to-sequence Artificial Neural Network (ANN) with Connectionist Temporal Classification (CTC), which maps a given audio file to a series of graphemes (characters) e.g. the english alphabet a-z.
tools.py contains a function which converts a raw audio file of samplerate 16kHz to a series of mel-spaced log-filterbanks features, using a stride of 10ms and window size 20ms.
The model consists of 3 linear layers followed by a recurrent network with options for bidirectionality as well as cell type e.g. GRU, LSTM or RNN, following this is another linear layer and finally a linear projection to form the unnormalised log probabilities (logits) for the CTC input.