Speech/non-speech detection in Malay language spontaneous speech

The goal of this work is to discriminate speech and non-speech segments in Malay language spontaneous speech as speech/non-speech detection is important in many speech processing applications. Inaccurate sentence boundaries are a major cause of errors in automatic speech recognition and a preprocessing stage that segments the speech signal into periods of speech and non-speech is invaluable in improving the recognition accuracy. We proposed a combination of three audio features that is energy, zero crossing rate (ZCR) and fundamental frequency (F0) for the speech/non-speech detection as each feature has unique properties to differentiate speech and non-speech segments. Experiments are conducted on one-hour Malay language spontaneous speech consisting of more than 20,000 speech/non-speech segments. An accuracy evaluation reveals that the proposed method achieved 97.8% accuracy rate. Non-speech segments will further be used as candidates of sentence boundary in our next experiment.