Automatic Recognition of Isolated Spoken Words with New Features

Abstract

Three new features are proposed for isolated spoken word recognition by computer. The features may be extracted from digital filterbank analysis. They are essentially approximation to subjective loudness, spectral balance and first and second formant ratio. The features have been tested on a set of English words spoken by different persons of British accent and the results are given. A dynamic programming algorithm is used to match templates of single word utterances with the trial utterances, for each of the features separately. The results indicate that the proposed features are quite useful in isolated word recognition.

Indexing terms:

• Speech recognition
• Digital filterbank
• Feature extraction
• Dynamic programming

Additional information

Notes on contributors

B.B. Chaudhuri

B.B. Chaudhuri: For biographical sketch of the author, please refer to JIETE, vol 26, p 28, Jan 1980.