{"id":669,"date":"2019-12-24T09:21:56","date_gmt":"2019-12-24T08:21:56","guid":{"rendered":"http:\/\/pepason.fr\/?p=669"},"modified":"2020-09-22T17:17:14","modified_gmt":"2020-09-22T15:17:14","slug":"deeplearning-et-soundscapes-par-q-a","status":"publish","type":"post","link":"https:\/\/pepason.fr\/?p=669","title":{"rendered":"“Deeplearning” et “Soundscapes” par Q.A"},"content":{"rendered":"\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t

Cr\u00e9er des paysages sonore... Un projet d'informatique de Quentin Aristote<\/h1>\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t
\n\t\t\t\t

Comment d\u00e9finir LE paysage sonore d\u2019un espace ? En effet, comment d\u00e9finir ce qui change en permanence ?\u00a0 Comment le reproduire avec suffisamment de pr\u00e9cision pour que notre oreille le reconnaisse et dise : “Ah oui, ce son vient de l\u00e0 !” ?…<\/em><\/p>

Pour r\u00e9pondre \u00e0 cette question, Quentin Aristote, \u00e9l\u00e8ve en Master 1 d’Informatique \u00e0 l’Ecole Normale Sup\u00e9rieure s’associe \u00e0 PePaSon et propose de coder un algorithme de “deeplearning<\/a>” pour\u00a0apprendre<\/strong> \u00e0 partir d’enregistrements r\u00e9elles \u00e0 reconstruire<\/strong> des paysages sonores… irr\u00e9els mais similaires.\u00a0<\/p>\t\t\t\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t\n\t\t\t\t\t\t\n\t\t\t\t\t\t\t\n\t\t\t\t<\/i>\t\t\t<\/span>\n\t\t\t\t\t\tT\u00e9l\u00e9charger le code \"open-source\" sur GitHub<\/span>\n\t\t<\/span>\n\t\t\t\t\t<\/a>\n\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t
Extrait du spectrogramme produit par l'algorithme \u00e0 la fin du projet. Ca ne marche pas encore... Mais c'est une belle tentative loin d'\u00eatre al\u00e9atoire !<\/figcaption>\n\t\t\t\t\t\t\t\t\t\t<\/figure>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t\n\t\t\t\t\t\t<\/span>\n\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t

Description du projet (Q.A) - Abstract<\/h4>\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t
\n\t\t\t\t
\n
\n
Version fran\u00e7aise :<\/strong><\/h6>\n

Le but du projet est d\u2019\u00e9crire un code informatique capable de g\u00e9n\u00e9rer des paysages sonores<\/em>\u2026. Une mani\u00e8re de la faire est de faire appel \u00e0 une forme particuli\u00e8re d\u2019algorithme appel\u00e9e \u201cr\u00e9seau de neurone<\/a>\u201c. Ce dernier est capable, apr\u00e8s avoir travaill\u00e9 \u00e0 partir d\u2019une base de donn\u00e9es (i<\/em>.e. <\/em>de paysages sonores pr\u00e9-existants), d\u2019apprendre \u201ccomment\u201d est fait un paysage sonore et, une fois bien entrain\u00e9e, devrait pouvoir produire des sons similaires \u00e0 de v\u00e9ritables paysages sonores, totalement nouveaux.<\/p>\n

Il existe de nombreuses mani\u00e8res de faire ceci\u2026 De nombreux r\u00e9seaux de neurones existents d\u00e9j\u00e0, travaillant avec des fichiers sonores et une multitude de bases de donn\u00e9es sont disponibles\u2026. La question reste \u00e0 savoir : \u201d Quel type de paysage sonore veut-on g\u00e9n\u00e9rer ? Sur quelle dur\u00e9e et admettant quelles variations ? \u201c.<\/em> Ces param\u00e8tres sont d\u00e9terminants dans le choix de la m\u00e9thode \u00e0 utiliser.<\/p>\n

Pour commencer nous travaillerons \u00e0 partir de la riche sonoth\u00e8que<\/a> du Mus\u00e9um National d\u2019Histoire Naturelle de Paris (France) sur des enregistrements de 30s libres de droits<\/em>. Nous utiliserons un \u201cAuto-Encoder<\/em><\/a>\u201d, i.e.<\/em> un reseau de neurone compose de deux sous-r\u00e9seaux, l\u2019un travaillant \u00e0 extraire les informations les plus simples possible de chaque paysage sonore (enregistrement) sous forme d\u2019un court vecteur, l\u2019autre travaillant \u00e0 recomposer un paysage sonore \u00e0 partir de la multitude de petits vecteurs extraits par le premier.<\/p>\n

Th\u00e9oriquement nous pourrions ainsi obtenir un paysage sonore \u201cressemblant\u201d \u00e0 ceux sur lesquels s\u2019est entrain\u00e9 l\u2019ordinateur.<\/p>\n

Nous n\u2019avons n\u00e9anmoins qu\u2019un faible contr\u00f4le sur le produit fini de l\u2019algorithme et l\u2019ajout de contraintes se fera dans un second temps. Cela reste th\u00e9oriquement complexe et n\u00e9cessite des informations tierces parfois indisponibles (commentaires, meta-donn\u00e9es\u2026).<\/p>\n

Un travail futur sera d\u2019appliquer ce r\u00e9seau \u00e0 un environnement unique afin d\u2019en extraire et g\u00e9n\u00e9rer les principaux aspects. Nous pourrons aussi travailler avec des donn\u00e9es labelis\u00e9es et comment\u00e9es afin d\u2019extraire et de produire des informations qualitatives sur les paysages sonores.<\/p>\n

Mais gardons en t\u00eate\u2026 Les r\u00e9seaux de neurones sont encore mal compris et \u00e9chappent aux math\u00e9matiques contemporains les rendant parfois difficiles \u00e0 utiliser\u2026<\/p>


<\/p>\n<\/div>\n<\/div>\n

\n
\n
English version :<\/strong><\/em><\/h6>\n

The goal of the project is to design a neural network whose aim is to generate soundscapes from scratch : although we would need a database of pre-existing soundscapes to train the neural network, once it is fully trained it should be able to generate new random soundscapes that sound like real soundscapes.<\/em><\/p>\n

They are a lot of ways to do this. Many neural network architectures working well with sound recordings currently exist and have already been extensively studied. Similarly, there are many ways to choose the training database : it could consist in very different soundscapes as well as in many soundscapes recorded in the same environment, for example. Ultimately, all these choices need to be made with what type of soundscapes we want to generate. Do we want them to be diverse or variations on the same soundscape ? Do we want them to last a few seconds or a few hours ?<\/em><\/p>\n

For starters, I will be working with the Paris Natural History Museum soundbank, that provides open-access 30 seconds samples of many different soundscapes. I will thus try to design an Auto Encoder that is able to generate 30 seconds long soundscapes that may be very diverse. An Auto Encoder is a special kind of neural network that is actually composed of two sub-networks. The first one, the Encoder, tries to reduce the dimension of the data, i.e. extract tje main features, and thus outputs a small-dimensional vector for each soundscape. The second one, the Decoder, takes those small dimensional vectors and tries to reconstruct the original soundscapes from them. Once this network is trained, we can try to generate new soundscapes by taking a random vector and feeding it to the Decoder, hoping that the small-dimensional vector space in between the Encoder and the Decoder is a faithful representation of the space of soundscapes.<\/em><\/p>\n

In practice, this means that we don\u2019t get to choose what kind of soundscape the Decoder will generate, but by changing the Auto Encoder\u2019s architecture a little bit it is theoretically possible to add a feature selector. Of course, as long as we don\u2019t have any metadata on our recordings, it won\u2019t be possible to choose which features this feature selector would allow us to select.<\/em><\/p>\n

Once this first network works well, it should be possible to train it with other kind of soundbanks. For example, we\u2019d like to focus on soundcapes recorded in a single environment. It should also be possible to work with pre-labeled data (if we get some) to control the kind of soundscapes we generate and increase the training speed.<\/em><\/p>\n

All of this is of course only theoretical yet, and we have no guarantee it should work except the fact that it did in the past for other kinds of data : neural networks aren\u2019t very well understood mathematically.<\/em><\/p>\n<\/div>\n<\/div>\t\t\t\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"

Quentin Aristote s’est associ\u00e9 \u00e0 PePaSon pour son projet de Master en Informatique. L’objectif : coder un programme capable de cr\u00e9er ab initio un paysage sonore aux propri\u00e9t\u00e9s similaires \u00e0 celles d’un environnement donn\u00e9 et pr\u00e9-enregistr\u00e9. <\/p>\n","protected":false},"author":1,"featured_media":671,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"cybocfi_hide_featured_image":""},"categories":[6,5],"tags":[],"post_mailing_queue_ids":[],"_links":{"self":[{"href":"https:\/\/pepason.fr\/index.php?rest_route=\/wp\/v2\/posts\/669"}],"collection":[{"href":"https:\/\/pepason.fr\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/pepason.fr\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/pepason.fr\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/pepason.fr\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=669"}],"version-history":[{"count":6,"href":"https:\/\/pepason.fr\/index.php?rest_route=\/wp\/v2\/posts\/669\/revisions"}],"predecessor-version":[{"id":699,"href":"https:\/\/pepason.fr\/index.php?rest_route=\/wp\/v2\/posts\/669\/revisions\/699"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/pepason.fr\/index.php?rest_route=\/wp\/v2\/media\/671"}],"wp:attachment":[{"href":"https:\/\/pepason.fr\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=669"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/pepason.fr\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=669"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/pepason.fr\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=669"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}