Trained on: Mozilla Corpus, Personal Recordings, and LibriSpeech — 2900 hrs of audio data. Training Script and Experiment Results available here