PhD thesis proposal

Incremental sequence-to-sequence mapping for

speech generation using deep neural networks

June 17, 2019

1 Context and objectives

In recent years, deep neural networks have been widely used to address sequence-

to-sequence (S2S) learning. S2S models can solve many tasks where source

and target sequences have di
erent lengths such as: automatic speech recog-

nition, machine translation, speech translation, text-to-speech synthesis, etc.

Recurrent, convolutional and transformer architectures, coupled with attention

models, have shown their ability to capture and model complex temporal de-

pendencies between a source and a target sequence of multidimensional discrete

and/or continuous data. Importantly, end-to-end training alleviates the need

to previously extract handcrafted features from the data by learning hierarchi-

cal representations directly from raw data (e.g. character string, video, speech

waveform, etc.).

The most common models are composed of an encoder that reads the full in-

put sequence (i.e. from its beginning to its end) before the decoder produces the

corresponding output sequence. This implies a latency equals to the length of

the input sequence. In particular, for a text-to-speech (TTS) system, the speech

waveform is usually synthesized from a complete text utterance (e.g. a sequence

of words with explicit begin/end-of-utterance markers). Such approach cannot

be used in a truly interactive scenario, in particular by a speech-handicapped

person to communicate orally'. Indeed, the interlocutor has to wait for the

complete utterance to be typed before being able to listen to the synthetic voice,

hence limiting the dynamics and naturalness of the interaction.

The goal of this project is to develop a general methodology for incremental

sequence-to-sequence mapping, with application to interactive speech technolo-

gies. It will require the development of end-to-end classication and regression

neural models able to deliver chunks of output data on-the-y, from only a par-

tial observation of input data. The goal is to learn an ecient policy that leads

to an optimal trade-o
between (variable) latency and accuracy of the decoding

process. Possible strategies to decode the output data as soon as possible in-

clude: (i) Predicting online he future' of the output sequence from he past

1

and present' of the input sequence, with an acceptable tolerance to possible er-

rors, or (2) learn automatically from the data an optimal waiting policy' that

prevents the model to output data when the uncertainty is too high. The devel-

oped methodology will be applied to address two speech processing problems:

(i) Incremental Text-to-Speech synthesis in which speech is synthesized while

the user is typing the text (possibly with a variable latency), and (ii) Incremen-

tal speech enhancement/inpainting in which portions of the speech signal are

unintelligible because of sudden noise or speech production disorders, and must

be replaced on-the-y with reconstructed portions.

2 Work plan

The proposed working plan is the following :

 Bibliographic work on S2S neural models, in the context of speech recogni-

tion, speech synthesis, and machine translation as well as their incremental

(low-latency) variations

 Investigating new architectures, losses, and training strategies toward in-

cremental S2S models.

 Implementing and evaluating the proposed techniques in the context of

end-to-end neural TTS systems (the baseline system may be a neural

TTS trained with past information/left-context only).

 Implementing and evaluating the proposed techniques in the context of

speech enhancement/inpainting, rst on simulated noisy speech and then

on pathological speech.

3 Requirements

We are looking for an outstanding and highly motivated PhD candidate to work

on this subject. Following requirements are mandatory:

 Engineering degree and/or a Master's degree in Computer Science, Signal

Processing or Applied Mathematics.

 Solid skills in Machine Learning. General knowledge in natural language

processing and/or speech processing.

 Excellent programming skills (mostly in Python and deep learning frame-

works).

 Good oral and written communication in English.

 Ability to work autonomously and in collaboration with supervisors and

other team members.

2

4 Work context

Grenoble Alpes Univ. o
ers an excellent research environment with ample com-

puting facilities, as well as remarkable surroundings to explore over the week-

ends. The PhD project will be funded by the Grenoble Articial Intelligence

Institute (MIAI). The PhD candidate will work both at GIPSA-lab (CRISSP

team) and LIG-lab (GETALP team). The duration of the PhD is 3 years. The

salary is between 1770 and 2100 euros gross per month (depending on comple-

mentary activity or not).

5 How to apply?

Applications should include a detailed CV; a copy of their last diploma; at least

two references (people likely to be contacted); a cover letter of one page; a one-

page summary of the Master thesis; the two last transcripts of notes (Master or

engineering school). Applications should be sent to thomas.hueber@gipsa-lab.fr,

laurent.girin@gipsa-lab.fr and laurent.besacier@imag.fr. Applications will be

evaluated as they are received: the position is open until it is lled, with deadline

on July 10th, 2019.