Abstract: An algorithm of speech detection in the time-frequency domain based on matching pur-suit decomposition is proposed.With matching pursuits,decomposed time-frequency energy of the signal has a clear distribution in the time-frequency domain.Using this characteristic,the endpoint can be differentiated by setting a threshold value.Compared with the conventional method that uses zero crossing ratio to detect surd and short-time energy to detect sonant,although the real-time capability is inferior to the conventional method,the endpoint can be detected accurately.The acc-uracy is much better than the conventional method in an environment without noise.