{"id":3998,"date":"2025-12-19T13:12:40","date_gmt":"2025-12-19T12:12:40","guid":{"rendered":"https:\/\/aholab.ehu.eus\/aholab\/?p=3998"},"modified":"2025-12-26T13:29:38","modified_gmt":"2025-12-26T12:29:38","slug":"libribrain-2025-primer-puesto","status":"publish","type":"post","link":"https:\/\/aholab.ehu.eus\/aholab\/libribrain-2025-primer-puesto\/","title":{"rendered":"LibriBrain 2025: 1st position"},"content":{"rendered":"\n

This year, we took part in the LibriBrain Competition 2025<\/strong>, an international challenge presented at NeurIPS 2025<\/strong> aimed at advancing the decoding of language from non-invasive brain signals using the large-scale LibriBrain dataset. The competition seeks to foster progress in brain\u2013computer interfaces, with the long-term goal of restoring communication abilities in individuals with speech impairments and enabling novel forms of human\u2013machine interaction based on neural data.<\/p>\n\n\n\n

Our system, neural2speech<\/strong>, achieved first place in the Phoneme Classification Standard Track<\/em><\/strong>. This track focuses on the prediction of phoneme classes directly from MEG (magnetoencephalography) recordings, under a constrained setting where only the official training data can be used, making robustness and generalization key challenges.<\/p>\n\n\n\n

In our paper, \u201cMEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification<\/a>\u201d<\/em>, we describe the technical approach behind our solution. We adapt a Conformer architecture<\/strong>, originally designed for automatic speech recognition, to operate directly on raw MEG signals from 306 sensors, effectively capturing both temporal dependencies and spectral characteristics of neural activity. Our method incorporates instance-level normalization<\/strong> to reduce distribution shifts across data splits, a dynamic chunk-averaging data loader<\/strong> to improve phoneme classification performance, and class-balancing strategies<\/strong> based on inverse square-root frequency weighting to address class imbalance.<\/p>\n\n\n\n

These design choices result in a robust and competitive system, allowing neural2speech to stand out among the submitted solutions and demonstrating a meaningful step forward at the intersection of neural signal processing and speech technology.<\/p>\n","protected":false},"excerpt":{"rendered":"

This year, we took part in the LibriBrain Competition 2025, an international challenge presented at NeurIPS 2025 aimed at advancing the decoding of language from non-invasive brain signals using the large-scale LibriBrain dataset. The competition seeks to foster progress in brain\u2013computer interfaces, with the long-term goal of restoring communication abilities in individuals with speech impairments…<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_es_post_content":"\n

Este a\u00f1o hemos participado en la\u00a0LibriBrain Competition 2025<\/strong>, un desaf\u00edo internacional presentado en\u00a0NeurIPS 2025<\/strong>\u00a0que impulsa la investigaci\u00f3n de decodificaci\u00f3n de lenguaje a partir de se\u00f1ales cerebrales no invasivas, utilizando el extenso conjunto de datos LibriBrain. El objetivo de la competici\u00f3n es fomentar avances significativos en interfaces cerebro-ordenador que puedan, en el futuro, ayudar a restaurar la comunicaci\u00f3n en personas con d\u00e9ficits del habla y abrir nuevas v\u00edas para la interacci\u00f3n entre humanos y m\u00e1quinas a trav\u00e9s de se\u00f1ales neurales.\u00a0<\/p>\n\n\n\n

Nuestro sistema\u00a0neural2speech<\/strong>\u00a0logr\u00f3 el\u00a0primer puesto en el\u00a0Phoneme Classification Standard Track<\/em><\/strong>, una de las dos pistas del desaf\u00edo de clasificaci\u00f3n de fonemas que exige a los participantes desarrollar modelos capaces de predecir fonemas a partir de datos de MEG (magnetoencefalograf\u00eda), utilizando \u00fanicamente el conjunto de entrenamiento oficial sin datos externos.\u00a0<\/p>\n\n\n\n

En el paper\u00a0\u201cMEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification<\/a>\u201d<\/em>\u00a0describimos la\u00a0arquitectura y los m\u00e9todos clave<\/strong>\u00a0detr\u00e1s de nuestro enfoque: adaptamos un\u00a0modelo Conformer<\/strong>, originalmente desarrollado para tareas de reconocimiento autom\u00e1tico de voz (ASR), a se\u00f1ales MEG de 306 canales para capturar caracter\u00edsticas temporales y espectrales relevantes. Implementamos t\u00e9cnicas como\u00a0normalizaci\u00f3n a nivel de ejemplo<\/strong>\u00a0para mitigar cambios de distribuci\u00f3n entre particiones de datos y un\u00a0cargador din\u00e1mico de agrupamiento<\/strong>\u00a0para mejorar la clasificaci\u00f3n de fonemas en promedios de muestras, junto con un esquema de ponderaci\u00f3n de clases basado en la ra\u00edz inversa del n\u00famero de ejemplos para manejar el desequilibrio de clases.\u00a0<\/p>\n\n\n\n

Gracias a estas innovaciones, el modelo alcanza resultados robustos en el Standard Track de clasificaci\u00f3n de fonemas, destac\u00e1ndose entre las soluciones presentadas y consolidando un avance significativo en la intersecci\u00f3n entre procesamiento del lenguaje y neurociencia.<\/p>\n","_es_post_name":"","_es_post_excerpt":"","_es_post_title":"LibriBrain 2025: Primer Puesto","_eu_post_content":"\n

Aurten, LibriBrain Competition 2025 nazioarteko erronkan parte hartu dugu. NeurIPS 2025en aurkeztu da, eta garuneko seinale ez-inbaditzaileetatik abiatuz lengoaia deskodetzeko ikerketa bultzatzen du, LibriBrain datu-multzo zabala erabiliz. Lehiaketaren helburua ondorengoa da: garun-ordenagailu interfazeetan aurrerapen esanguratsuak sustatzea, etorkizunean hizketa-gabeziak dituzten pertsonen komunikazioa berrezartzen laguntzeko eta seinale neuralen bidez gizakien eta makinen arteko interakziorako bide berriak irekitzeko.

Gure neural2speech sistemak lehen postua lortu zuen Phoneme Classification Standard Track delakoan, hau da, fonemak sailkatzeko erronkaren bi pistetako batean; izan ere, parte-hartzaileek MEG (magnetoentzefalografia) datuetatik abiatuta fonemak aurreikusteko gai diren ereduak garatu behar dituzte, kanpoko daturik gabeko entrenamendu ofizialaren multzoa soilik erabiliz.

\"MEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification<\/a>\" paperean arkitektura eta metodo gakoak deskribatzen ditugu gure ikuspegiaren atzean: Conformer eredu bat, jatorrian ahotsaren ezagutza automatikoko zereginetarako garatua (ASR), 306 kanaleko MEG seinaleetara egokitzen dugu, denborazko eta espektrozko ezaugarri garrantzitsuak atzemateko. Adibide mailako normalizazio-teknikak inplementatzen ditugu, datu-partizioen eta taldekatzeko kargagailu dinamiko baten arteko banaketa-aldaketak arintzeko, laginen batez bestekoetan fonemen sailkapena hobetzeko, klaseen ponderazio-eskema batekin batera, klaseen desoreka maneiatzeko adibide-kopuruaren alderantzizko erroan oinarritua.

Berrikuntza horiei esker, ereduak emaitza sendoak lortzen ditu fonemak sailkatzeko Standard Track delakoan, aurkeztutako soluzioen artean nabarmenduz eta hizkuntzaren prozesamenduaren eta neurozientziaren arteko intersekzioan aurrerapen esanguratsua sendotuz.<\/p>\n","_eu_post_name":"","_eu_post_excerpt":"","_eu_post_title":"LibriBrain 2025: Lehenengoak","_en_post_content":"\n

This year, we took part in the LibriBrain Competition 2025<\/strong>, an international challenge presented at NeurIPS 2025<\/strong> aimed at advancing the decoding of language from non-invasive brain signals using the large-scale LibriBrain dataset. The competition seeks to foster progress in brain\u2013computer interfaces, with the long-term goal of restoring communication abilities in individuals with speech impairments and enabling novel forms of human\u2013machine interaction based on neural data.<\/p>\n\n\n\n

Our system, neural2speech<\/strong>, achieved first place in the Phoneme Classification Standard Track<\/em><\/strong>. This track focuses on the prediction of phoneme classes directly from MEG (magnetoencephalography) recordings, under a constrained setting where only the official training data can be used, making robustness and generalization key challenges.<\/p>\n\n\n\n

In our paper, \u201cMEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification<\/a>\u201d<\/em>, we describe the technical approach behind our solution. We adapt a Conformer architecture<\/strong>, originally designed for automatic speech recognition, to operate directly on raw MEG signals from 306 sensors, effectively capturing both temporal dependencies and spectral characteristics of neural activity. Our method incorporates instance-level normalization<\/strong> to reduce distribution shifts across data splits, a dynamic chunk-averaging data loader<\/strong> to improve phoneme classification performance, and class-balancing strategies<\/strong> based on inverse square-root frequency weighting to address class imbalance.<\/p>\n\n\n\n

These design choices result in a robust and competitive system, allowing neural2speech to stand out among the submitted solutions and demonstrating a meaningful step forward at the intersection of neural signal processing and speech technology.<\/p>\n","_en_post_name":"libribrain-2025-primer-puesto","_en_post_excerpt":"","_en_post_title":"LibriBrain 2025: 1st position","edit_language":"en","footnotes":""},"categories":[171,58,174],"tags":[],"class_list":["post-3998","post","type-post","status-publish","format-standard","hentry","category-beriak-ekitaldiak","category-noticias-y-eventos","category-noticias-y-eventos-es"],"_links":{"self":[{"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/posts\/3998","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/comments?post=3998"}],"version-history":[{"count":3,"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/posts\/3998\/revisions"}],"predecessor-version":[{"id":4001,"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/posts\/3998\/revisions\/4001"}],"wp:attachment":[{"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/media?parent=3998"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/categories?post=3998"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aholab.ehu.eus\/aholab\/wp-json\/wp\/v2\/tags?post=3998"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}