Niger-Mali Audio Collection [DATA][PAPER]

This unannotated audio collection corresponds to 671 hours of radio broadcasts in five languages: French from Niger, Fulfulde, Hausa, Tamasheq and Zarma.

Tamasheq-French Parallel Corpus [DATA][PAPER]

19 hours of Tamasheq audio from radio broadcasts in Niger, aligned to French text translations. Made available in the context of IWSLT 2022.

Mboshi-French Parallel Speech Corpus [DATA] [PAPER]

5,130 sentences long data set from a true language documentation case.  Mboshi (or Embosi) is a dialect spoken in Congo-Brazzavile. Multilingual extension available here.

Griko-Italian Parallel Speech Corpus [DATA] [PAPER]

330 sentences long data set from a true language documentation case.  Griko is a dialect spoken in south Italy, being a mix of Greek and Italian. 

MaSS Multilingual Speech-to-Speech Corpus [DATA][PAPER]

A aligned speech extension from the CMU data set (monolingual alignments from the Bible).  The results is a 8,130 sentences long speech parallel dataset.

The eight covered languages are: English, Spanish, French, Hungarian, Romanian, Basque, Russian and Finnish. 

mHuBERT-147 project:

• mHuBERT-147 pre-trained models [LINK]

• The fairseq fork for training [LINK]

• Pre-processing and clustering scripts [LINK]

• HUTTER: a mHuBERT-147 CommonVoice Prototype [LINK]

Multilingual DistilWhisper:

• Language Experts collection [LINK]

• Code for training and inference [LINK]

LeBenchmark and LeBenchmark 2.0 pre-trained models:

• LeBenchmark French models (13 models, from 1K to 14K hours) [LINK]

• LeBenchmark pre-processing and training scripts [LINK]

IWSLT 2023 - Multilingual Speech Translation:

• Recipe to the winning submission [LINK]

IWSLT 2022 - French to Tamasheq Direct Speech Translation:

• SpeechBrain recipe for low-resource speech translation [LINK]

• wav2vec 2.0 model trained on Tamasheq speech (243 hours) [LINK]

• wav2vec 2.0 model trained on the Niger-Mali audio collection (658 hours) [LINK]

2024:

• Seminar at CoML research group, ENS, Paris "mHuBERT-147: A Compact and Powerful Multilingual Speech Foundation Model" [SLIDES]

• UTTER DAYS 2024 "mHuBERT-147: A Compact Multilingual HuBERT Model" (Subset of slidedeck above)

2023:

• UTTER DAYS 2023 "NAVER LABS Europe's Multilingual Speech Translation Systems for the IWSLT 2023 Low-Resource Track" [SLIDES]

2022:

• Seminar at GIPSA-LAB "A brief (low-resource) speech processing adventure in the realm of self-supervised models." [SLIDES] (24/11/22)

2021:

• Seminar "Attention-based Unsupervised Word Segmentation from Speech" [SLIDES] (distilled version of the PhD defense). Invited seminar at: 

  • • NAVER Labs Europe (07/04), 

    • Sheffield University (24/06), 

    • Avignon University (02/07),

    • TALEP, LIS - Marseille (30/08),

    • ISCA SIGML special group (29/11),

    • CoML, ENS Paris (14/12).

• PhD defense "Models and Resources for Attention-based Unsupervised Word Segmentation" [SLIDES] (09/07/2021)

2018: 

• LIG PhD Day [SLIDES] [POSTER]

• LICIA Workshop [SLIDES]

2017: 

• Masters Defense "Unsupervised Word Discovery Using Attentional Encoder-Decoder Models" [SLIDES]

2023:

• Invited panelist for Festival'IA 2023: "IA et multilinguisme : enjeux et état des lieux". 

• Invited panelist for RJCP 2023: "poursuites après le doctorat".

2022:

• Invited panelist for SIGUL 2022: towards building technologies for low-resource languages while respecting the sovereignty of the communities.