Provided are device, method, and medium for generating an audio fingerprint and retrieving audio data. The device for generating an audio fingerprint includes: a coefficient extracting section partially decoding audio data in a compression area and extracting MDCT (Modified Discrete Cosine Transform) coefficients; a coefficient selecting section selecting an MDCT coefficient robust to noises from the extracted MDCT coefficients; a modulation spectrum generating section transforming the selected MDCT coefficient by the use of a Fourier transform method and generating a modulation spectrum; and a bit conversion section quantizing the generated modulation spectrum and generating an audio fingerprint. As a result, it is possible to accurately and rapidly retrieve the audio data recorded in a variety of environments. Since elements based on MP3 are used, it is possible to apply to MP3 applications in various manners. In addition, it is possible to apply to classification of audio data such as classification of music moods and classification of music genres and various other fields such as extraction of a specific event from moving images of sports.