As one of the key methods to extract content semantics and structure from audio, automatic audio classification, especially for a speech and a music, is valuable for content-based audio retrieval, video summary and retrieval, and spoken document retrieval, etc. Because hidden Markov model (HMM) can well model audio signal抯 time statistical properties, a left-right discrete HMM is proposed to classify a speech, a music and their mixed audio. The experimental results show that HMM is excellent for audio classification accuracy is up to 90.28％.