Wav2Letter Convolutional ASR

The 2016 wav2letter paper by Facebook AI Research researchers combines a convolutional acoustic network with graph decoding for speech recognition. It predicts letters or graphemes from speech inputs and does not require an intermediate phonetic transcription for the stated training approach. The study evaluates inputs including MFCC features, power-spectrum features, and raw waveform. Calling it exclusively a raw-waveform system would therefore omit reported variants. Its acoustic architecture avoids recurrent network connections, while a complete decoder can still perform sequential search. The original implementation used Torch7, with criterion and decoder components implemented in C. This should be distinguished from the later wav2letter++ framework, which its paper describes as written in C++ using the ArrayFire tensor library. The original automatic segmentation criterion, or ASG, differs from CTC by omitting the blank symbol and including transition scores between letters. Both seek to train from sequence annotations without requiring manual labels at every audio frame, although their definitions differ. The authors report results on LibriSpeech under specific input and decoding choices; the architecture name alone does not guarantee a latency or error-rate advantage for every workload. Flashlight is a related C++ machine-learning library whose documented applications include ASR formerly associated with the wav2letter project. Convolutional stacks also have a receptive-field geometry worth separating from framework speed. In a stated illustration of two length-three, stride-one, dilation-one temporal layers with no padding or intervening operation, one second-layer output depends on five original input positions. This is a simplified geometry calculation, not a measurement of every wav2letter model.