Projects audio files that contains one word of speech into a hyper-dimension space just like Word2Vec. Uses "Force Aligment" to split audio into words (which requires text). Pad the audio segments with zeros, do MFCC, feed into encoder-decoder which uses RMSE. They also add noise to the signal and make the network denoise it. LibriSpeech 500 hour of audio. Not sure how it can incorporated in an ASR or TTS systems. The audio file has to be paired with a text otherwise Speech2Vec cannot split the audio file into words using "Forced Alignment" method. It is used to query if the spoken word is similar to an existing word in the corpus.
BPE data compression tool that combines most frequent pair of bytes with one. It works well with Named Entity, loadwords and morphologically complex words. Handles OOVs well and rare words. You can