A Neural Network Model for Phoneme Generation

Abstract

The paper discusses the possibility of phonemes generation based on a recurrent neural network model. In each phoneme a typical or elemental pattern can be identified that repeats itself with slight fluctuations along the signal length. This elemental pattern constitutes the training data for the recurrent neural network. After training, the network can generate three new periods of elemental patterns. In a repetitive loop the network can generate the entire phoneme signal. The model proved very simple and effective, and the generated phonemes gave the impression of a natural sound.

Phoneme Generation with Elman-type Neural Networks

“…The proposed approach continues the main idea introduced in some previous works [7,8] this way, the network was completely independent from the training data and predicted the continuation of the input signal with 365 new samples. The final phase of simulation involved the generation of elementals with slight variations of the signal shape.…”

“…The suggested model for phoneme synthesis consisted of the generation of the elemental patterns by a harmonic series and the repetition of that pattern in time with the signal parameters controlled by a chaotic source. In another work, we investigated the possibility to generate the phoneme elementals through recurrent neural networks [8]. Neural networks may offer a good alternative for the dynamic approach of speech synthesis due to their inherent capability of capturing complex nonlinear relations in data.…”

“…In general, the prediction performance is acceptable but only for one or a few instances in advance. In [8] we investigated the possibility of training a similar topology for the generation of three new periods of elemental pattern, which we considered to be a quite good long-run result. The phoneme sound was finally generated in a repetitive loop.…”