title={Speaker interpolation in HMM-based speech synthesis system},
author={Takayoshi Yoshimura and Takashi Masuko and Keiichi Tokuda and Takao Kobayashi and Tadashi Kitamura},
booktitle={EUROSPEECH},
year={1997}
}
@article{STRAIGHT,
title="Restructuring speech representations using a pitch-adaptive time–frequency smoothing and an instantaneous-frequency-based F0 extraction: Possible role of a repetitive structure in sounds1Speech files available. See http://www.elsevier.nl/locate/specom1",
abstract={{Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day}},
author={Lecun, Y. and Bottou, L. and Bengio, Y. and Haffner, P.},
@@ -73,9 +73,32 @@ To do when I'll have a good overview of the project. Try to answer:
\color{red}
Concise presentation of the problem
Methods before deep learning
*Note that layers will be explained in an upcoming section*
Wavenet
Preprocessing of text into phonemes?
SOTA ON MULTISPEAKER TTS:
First SPSS methods [2 - 20] of https://arxiv.org/pdf/1606.06061.pdf
\color{black}
Previous state of the art in TTS include hidden Markov models (HMM) based speech synthesis, which is a statistical parametric speech synthesis (SPSS) method. HMMs are trained to synthesize mel-frequency cepstral coefficients (MFCC) with energy, their delta and delta-delta coefficients \cite{TTSSOTA}. The result is passed through a vocoder\footnote{Specifically in TTS, some authors define a vocoder as a voice encoder that retrieves speech parameters to be used in synthesis. The more common definition however, is that of a function that generates a raw audio waveform from temporal features such as MFFC. This is the one we will use. \color{red} Review this \color{black}} such as MLSA \cite{MLSA}. The spectral parameters, pitch parameters and state durations of the model are conditioned on the phoneme contexts such that different contexts are clustered by a decision tree and a distribution is learned for each cluster \cite{HMMTTS}. It is thus possible to modify the voice generated by conditioning on a speaker or tuning these parameters with adaptation or interpolation techniques (e.g. \cite{HMMSpeakerInterpolation}), effectively making HMM-based speech synthesis a multispeaker TTS system. \color{red} Compare with concatenative \cite{SPSSDNN} ? \color{black}
