Automatic speech recognition (ASR) is the process of transcribing recorded speech into text (words).

Speech recognition systems are mostly trained on adult speech. Their performance on child speech is generally degraded, due to the acoustic differences between child and adult voices [Gelin, 2021]. However, recent systems are performing better and better, and there is hope that with adaptation they will continue to do so when applied to children’s speech, possibly in noisy environments such as a classroom.

During this internship, we would like to address the following points:

• Testing ready-to-use systems on a child speech test set: Whisper [Radford, 2022] and other models available in the SpeechBrain toolkit [Parcollet, 2021];

• Adapt the most promising system to child speech (transfer learning);

• Explore self-supervised systems as alternative, such as Wav2Vec2.0 [Schneider, 2019], HuBERT [Hsu, 2021], WavLM [Chen, 2022]. In this case, several options are considered:

  • Train/adapt only the classifier part of the model (fine tuning) on a child speech training set;

  • Train/adapt the complete model (pretraining and fine tuning) on child speech.

The performance of the systems will be analysed in function of the age of the young speakers.

This internship is part of a close collaboration between Lalilo and IRIT. You will be supervised by Dr. Lucile Gelin (Associate researcher at Lalilo and IRIT) and Dr. Thomas Pellegrini (Associate professor at IRIT). You can choose to be based in the Lalilo Paris office or in the IRIT Toulouse office. If working from Paris, you will have at least 2 trips (a few days) to Toulouse to meet the IRIT team and discover the research laboratory’s environment.

We are looking for a 5-6 months internship, starting in September-October.

About you

• Rigorous in your work

• Being proactive

• Autonomous

• Interested in research

• Strong desire to learn

Required skill

• Python

• Theoretical knowledge in machine learning

• Technical English

Appreciated skills

• Pytorch

• Theoretical knowledge in speech processing and/or signal processing

• Interested in innovation for education

Please submit your resume and optionally a cover letter explaining why you would like to do your internship with us.
Then the recruitment process will consist in two steps :

• Motivation interview

• Technical test (research paper review)