Text-to-audio-visual speech synthesis based on parameter generation from HMM

Abstract

We present a new self-organizing neural network which performs unsupervised learning and can be used for vector quantization. The main advantage over existing approaches, e.g., the Kohonen feature map, is the ability of the model to automatically find a suitable network structure and size. This is achieved through a controlled growth process which also includes occasional removal of units. The algorithm is evaluated on a database that includes 25 speakers each of them recorded in 12 diffrent sesions. The overall performance was 99.5%. That is, in 99.5% of the trials, the right speaker was correctly accepted and the impostor speaker correctly rejected.