A Neural Model to Predict Parameters for a Generalized Command Response Model of Intonation

Abstract

The Generalised Command Response (GCR) model is a timelocal model of intonation that has been shown to lend itself to (cross-language) transfer of emphasis. In order to generalise the model to longer prosodic sequences, we show that it can be driven by a recurrent neural network emulating a spiking neural network. We show that a loss function for error backpropagation can be formulated analogously to that of the Spike Pattern Association Neuron (SPAN) method for spiking networks. The resulting system is able to generate prosody comparable to a state-of-the-art deep neural network implementation, but potentially retaining the transfer capabilities of the GCR model.

Synthetic Speech References for Automatic Pathological Speech Intelligibility Assessment

“…To this end, we use a DNN-based state-of-the-art Merlin TTS system in conjunction with the Festival front-end, two Bidirectional Long Short-Term Memory networks as duration and acoustic models, and the WORLD vocoder. For details on the TTS systems and the training procedure, the reader is referred to [23,24]. By training a TTS system for each speaker, we get 4 speaker-dependent TTS systems.…”

An End-to-end Network to Synthesize Intonation Using a Generalized Command Response Model

“…The E2E system is initialized with a pre-trained atom model, which uses the same topology and training as described in our previous work [3]. At first the E2E model is trained for 50 epochs (LR of 0.001), without the phrase bias, on LF0 from which the phrase contribution is removed.…”

“…We studied how a Recurrent Neural Network (RNN) can generate the command signals of the GCR model to generate intonation by emulating a spiking neural network [3]. The RNN predicts the position and amplitude of the command spikes for a given text, which are filtered by the GCR muscle models to generate the pitch contour.…”