Post-Doctorant a Brain-Computer Interface For The Decoding Of Language Production H/F - INRIA

A propos d'Inria

Inria est l'institut national de recherche dédié aux sciences et technologies du numérique. Il emploie 2600 personnes. Ses 215 équipes-projets agiles, en général communes avec des partenaires académiques, impliquent plus de 3900 scientifiques pour relever les défis du numérique, souvent à l'interface d'autres disciplines. L'institut fait appel à de nombreux talents dans plus d'une quarantaine de métiers différents. 900 personnels d'appui à la recherche et à l'innovation contribuent à faire émerger et grandir des projets scientifiques ou entrepreneuriaux qui impactent le monde. Inria travaille avec de nombreuses entreprises et a accompagné la création de plus de 200 start-up. L'institut s'eorce ainsi de répondre aux enjeux de la transformation numérique de la science, de la société et de l'économie.

Post-Doctorant F/H A Brain-Computer interface for the decoding of language production

Type de contrat : CDD

Niveau de diplôme exigé : Thèse ou équivalent

Fonction : Post-Doctorant

Niveau d'expérience souhaité : De 3 à 5 ans

A propos du centre ou de la direction fonctionnelle

Created in 2008, the Inria Saclay Center is located at the heart of the Paris-Saclay scientific and technological excellence cluster, which alone accounts for 15% of French research. Serving the development of the Université Paris-Saclay and the Institut Polytechnique de Paris, the Inria Saclay center employs 80 people in research support services and 500 scientists of 54 nationalities.

Benefiting from continuous growth, the center now has a total of 42 project-teams and two in the process of being created, including 21 jointly with the Institut Polytechnique de Paris and 16 with the Université Paris-Saclay.

Contexte et atouts du poste

Dans le cadre d'un partenariat (vous pouvez choisir entre)

- ANR Speakout

The expected outcome of SpeakOut is to enable three anarthric patients to restore spoken communication
at a rate comparable to natural speech (approximately five to six syllables per second) by decoding neural
activity recorded and transmitted via a fully implantable system. This will allow the patients to use their
speech neural decoders for daily communication in the long term. This outcome will be achieved by improving
the current technique in several ways, as described below.
The results of this project will inform the next clinical trial, particularly regarding the potential validity and
generalizability of the syllabic approach, as well as the tolerability and benefits that patients can expect from
a speech BCI.

The project is run in collaboration with teams in Paris downtoan (institut de l'audition, Lariboisière).

Mission confiée

A complete pipeline for online, real-time, and offline processing and analysis is being developed. It is based on
state-of-the-art frameworks designed to use neural time series to drive a BCI and decode speech as discussed
in Metzger et al. [2023], Willett et al. [2023]. After preprocessing, relevant features including different
frequency bands, such as low beta, broadband high-frequency activity, and cross-frequency coupling Metzger
et al. [2023], Proix et al. [2022] are extracted for use in the decoding stage. These features are decoded into
speech using an artificial neural network (ANN) architecture that performs three primary operations:
1. Speech detection: To detect speech attempts, we will use a long short-term memory (LSTM) archi-
tecture because it is well-suited for modeling long-term dependencies and for real-time inference with
temporally dynamic processes Hochreiter and Schmidhuber [1997]. In the original setup, the LSTM
(which may be bidirectional) will be trained using a connectionist temporal classification (CTC) loss
function, as this has been shown to improve alignment in the absence of clear timing information in
covert speech Metzger et al. [2023], Willett et al. [2023].
2. Speech classification: Each detected window of relevant neural activity (speech attempt) is passed to a
multi-layer ANN classifier that computes syllable probabilities. First, the classifier down-samples the
neural activity with a temporal convolution to create a higher-dimensional data representation. Then,
this representation is processed by bidirectional gated recurrent unit (GRU) layers and a fully connected
(dense) layer that projects the latent dimension from the final GRU layer to probability values across
the syllable classes.
3. Language modeling: We plan to leverage and adapt pre-trained speech/language models (e.g. Wav2vec2
and GTP-4) in conjunction with a e.g. Viterbi algorithm to label syllables and determine the most
likely syllable sequence, and infer the most probable word sequences given contextual information.
Risk mitigation. Despite the promising and exciting recent results demonstrating high-level performance
with a BCI in ALS Card et al. [2024], even with a large vocabulary, there is a risk that the decoding accuracy
will be insufficient for patients to choose the speech neural decoder as their preferred means of communication.
In this worst-case scenario, we would restrict the speech neural decoder to a corpus of the most ecologically
valuable vocabulary and sentences (e.g., the most common verbal interactions with caregivers or loved ones).
We also plan to combine features of invasive SND with those recorded from additional non-invasive EEG.
In that case, the non-invasive information would determine the pragmatics of the intended speech, such as
assertive versus directive speech Li and Negoita [2018]) and thus, narrow down the language model.

Principales activités

The projectwill begin with the analysis of publicly available data or data within the Speakout consortium,
such as SEEG or ECOG data from epileptic patients and possibly MEG data. The Speakout consortium
comprises researchers from the Institut de l'Audition at the Pasteur Institute and the APHP hospital in
Paris. The goal is to reduce the decoding error rate by using all possible means. The project will produce
experimental code in Python (PyTorch, etc.). In the long term, the code will be made available as an
open-source project to ensure scientific reproducibility.
Once the technology is sufficiently advanced, it will be used to create a closed-loop interface with an implanted
patient at APHP. In another line of research, assuming that data are available, we will consider deciphering
inner speech (as oposed to intended speech) Kunz et al. [2025].

Compétences

Previous experience in BCIs is a strong asset as it will make it easier for the candidate to understand the
concepts and tools involved. Knowledge of scientific computing in Python (NumPy, SciPy, and PyTorch) is re-
quired. The candidate will benefit from the Mind team's numerous developments, as well as its computational
facilities and expertise in various domains (machine learning, optimization, statistics, and neuroscience).

Avantages

- Restauration subventionnée
- Transports publics remboursés partiellement
- Congés: 7 semaines de congés annuels + 10 jours de RTT (base temps plein) + possibilité d'autorisations d'absence exceptionnelle (ex : enfants malades, déménagement)
- Possibilité de télétravail et aménagement du temps de travail
- Équipements professionnels à disposition (visioconférence, prêts de matériels informatiques, etc.)
- Prestations sociales, culturelles et sportives (Association de gestion des oeuvres sociales d'Inria)
- Accès à la formation professionnelle
- Sécurité sociale

Rémunération

2788 € gross / month