Section: New Results

Miscellaneous

In parallel with mainstream research activities, Linkmedia has a number of contribtions in other domains based on the expertise of the team members.

Bidirectional GRUs in spoken dialog

Participants : Christian Raymond, Vedran Vukotić.

Recurrent neural networks recently became a very popular choice for spoken language understanding (SLU) problems. They however represent a big family of different architectures that can furthermore be combined to form more complex neural networks. In [30], we compare different recurrent networks, such as simple recurrent neural networks, long short-term memory networks, gated memory units and their bidirectional versions, on the popular ATIS dataset and on MEDIA, a more complex French dataset. Additionally, we propose a novel method where information about the presence of relevant word classes in the dialog history is combined with a bidirectional gated recurrent unit (GRU).

Kernel principal components analysis with extreme learning machines

Participant : Christian Raymond.

Work in collaboration with M'Sila University, Algeria.

Nowadays, wind power and precise forecasting are of great importance for the development of modern electrical grids. In [26], we investigate a prediction system for time series based on kernel principal component analysis (KPCA) and extreme learning machine (ELM). Comparison with standard dimensionality reduction techniques show that the reduction of the original input space affects positively the prediction output.

Pronunciation adaptation for spontaneous speech synthesis

Work in collaboration with Gwénolé Lecorvé and Damien Lolive, IRISA, Rennes.

In [36], we present a new pronunciation adaptation method which adapts canonical pronunciations to a spontaneous style. This is a key task in text-to-speech as those pronunciation variants bring expressiveness to synthetic speech, thus enabling new potential applications. The strength of the method is to solely rely on linguistic features and to consider a probabilistic machine learning framework, namely conditional random fields, to produce the adapted pronunciations.