 |
Workshop on the Auditory Basis of Speech PerceptionKeele University, UK |
 |
|
Workshop on the Auditory Basis of Speech PerceptionKeele University, UK |
|
We propose a model able to identify superimposed patterns. Rather than a new connectionist scheme, this consists in controlling the recognition process performed by a 'classical' perceptron tuned by a regular learning stage. Hence, separation is based on a priori knowledge of the patterns, A key point of this model is the bifurcation process which is a creation of a second copy from a single-channel input. These two versions are processed in parallel. The dominant pattern is removed from the copy in order to get a second label. With the multilayer perceptron, we obtain remarkable results for the double-vowels (d-vowels) recognition task : when inputs are short-term (50 ms) warped place-coded spectra, double-recognition (d-recognition) score reaches 88% on a set of 6 synthetic vowels having various fundamental frequencies (f0). In addition, we briefly present a dual method based on linear discriminant analysis (LDA), working well with natural vowels, and allowing reconstruction of input signals.
Bibliographic reference. Varin, Laurent / Berthommier, Frédéric (1996): "Identification of concurrent sources with a perceptron-based model", In ABSP-1996, 222-225.