I wonder if the audio data from television programmes combined with subtitles would make a good training set. Here in the UK many of the subtitles are very good quality (presumably human transcribed), and would be easy to strip out non-language noises such as laughter (normally transcribed in brackets).