ISCA Archive - Speech recognition based on speech units (original) (raw)

Speech recognition based on speech units

G. Zanellato

In a classical quantization system, each vector is represented by the nearest centroid. But it is impossible then to distinguish two vectors belonging to the same class. In order to mitigate this disadvantage, we have taken into account the two nearest neighbours and also a "belonging degree" calculated from the distances between the vector and the two centroids. In the case of speaker independent speech recognition system, this "fuzzy" quantization gives better results.

For the recognition of large vocabularies, it is usual to perform a phonemic segmentation and then to achieve the matching on the obtained sequence. But it is very hard to set up a good modeling for the phonemes. Nevertheless, we can take into account elementary components of the phonemes, for which simplest models may be used. We investigate a set of 100 "acoustic units" that have been defined from the centroids of the fuzzy quantization. The results we obtained are very attractive.