Wav2Letter: Transforming Speech to Text with CNN for Automatic Speech Recognition

This paper introduces a straightforward approach to Automatic Speech Recognition(ASR) system that converts spoken language into written text, featuring a complete model that includes a convolutional neural network for the acoustic model. Unlike traditional methods that require aligning phonemes, our model is trained to directly output letters by transcribing speech. We've eliminated the need for force alignment. To achieve this, we've developed an automatic segmentation criterion for training. This criterion, which is as effective as the Connectionist Temporal Classification (CTC) method, allows us to train from sequence annotations without requiring explicit alignment. Our model produces competitive results in terms of word error rate when using the Librispeech corpus and Mel-Frequency Cepstral Coefficients (MFCC) features. Additionally, we've obtained promising results when working directly with raw waveform data.