F0 modeling using DNN for Arabic parametric speech synthesis

Imene Zangar (1) , Zied Mnasri (1) , Vincent Colotte (2) , Denis Jouvet (2)

1 ENIT - Ecole Nationale d'Ingénieurs de Tunis
2 MULTISPEECH - Speech Modeling for Facilitating Oral-Based Communication

Deep neural networks (DNN) are gaining increasing interest in speech processing applications, especially in text-to-speech synthesis. Actually state-of-the-art speech generation tools, like MERLIN and WAVENET are totally DNN-based. However, every language has to be modeled on its own using DNN. One of the key components of speech synthesis modules is the prosodic parameters generation module from contextual input features, and more particularly the fundamental frequency (F0) generation module. Actually F0 is responsible for intonation , that is why it should be accurately modeled to provide intelligible and natural speech. However, F0 modeling is highly dependent on the language. Therefore, language specific characteristics have to be taken into account. In this paper, we aim to model F0 for Arabic speech synthesis with feedforward and recurrent DNN, and using specific characteristic features for Arabic like vowel quantity and gemination, in order to improve the quality of Arabic parametric speech synthesis.

Mots clés: Fundamental frequency F0; Deep neural networks; Recurrent neural networks; Arabic parametric speech synthesis

Domaines: Traitement du signal et de l'image [eess.SP]; Intelligence artificielle [cs.AI]

Imene Zangar, Zied Mnasri, Vincent Colotte, Denis Jouvet. F0 modeling using DNN for Arabic parametric speech synthesis. INNSBDDL 2019 - INNS Big Data and Deep Learning, Apr 2019, Sestri Levante, Italy. ⟨hal-02177496⟩