videos in AudioSet
A list of about a million ten-second YouTube clips, each labelled with the sounds in it, that Google published in 2017. Labs use it to train and test general audio and music models.
Including 「知床旅情」Shiretoko ryojyoハーモニカと詩(ハーモニカ演奏), 「浜千鳥」Hamatidori ハーモニカと詩【唱歌】, 「大阪ラプソディー」Osaka Rhapsodyハーモニカと詩