PhD Candidate: Multi-task Learning in Diverse Speech Tasks
The Institute for Computing and Information Sciences (iCIS) at Radboud University is looking for a PhD candidate to study how multi-task learning can improve …
- Houtlaan, Nijmegen, Gelderland
- Tijdelijk contract / Tijdelijke opdracht
- Uren per week:
- 32 - 40 uur
- € 3061 - € 3061 per maand
The Institute for Computing and Information Sciences (iCIS) at Radboud University is looking for a PhD candidate to study how multi-task learning can improve the various tasks in speech technology, and what new deep learning architectures can be effective for this.
Traditionally, the various tasks in speech technology are treated as quite independent, leading to separate architectures and models. For instance, speech synthesis and speech recognition approaches tended to share very few components, perhaps only using common resources such as speech data and a pronunciation dictionary.
This changed with the introduction of deep learning approaches in speech technology. For instance, Google's Wavenet architecture has been used for both speech synthesis and speech recognition, and in modified form for voice transformation. Other approaches making a connection between synthesis and recognition are autoencoders and generative adversarial networks. Speech tasks that operate on the same type of data, for instance, speaker and speech recognition, are getting more integrated – the current state of the art in automatic speech recognition (ASR) uses i-vectors as side-information in the acoustic model for speech recognition, allowing the model to adapt to the characteristics of the speaker. Even though i-vectors as an embedding for ASR have been superseded by deep-learning-based x-vectors, the speaker and speech acoustic models are still separate entities.
One of the ways of sharing parts of a neural model is by employing multi-task learning, by sharing multiple layers at the front of the network, and using separate final layers for the different tasks. This approach has been used for multilingual ASR and joint speaker/speech recognition with various success. The goal of this research is to investigate more ways of sharing model capacity in a wider variety of tasks. There are many intrinsic and extrinsic factors that influence speech production and recording, and there are many databases designed to either recognise one factor (e.g. speaker identity) or be robust against other factors (e.g. recording channel, accent or language). Transferring robustness learned in one task to another task for which there is no explicit variability in the training material should be demonstrable by utilising clever deep-learning architectures, whereby we want to benefit maximally from new approaches researched in, for instance, computer vision.
You will pursue your PhD in a vibrant international research environment in the data science group at iCIS, with top-class professors and researchers in information retrieval and machine learning. At iCIS we value a diverse workforce. Female candidates are therefore particularly encouraged to apply.
- A Master’s degree in Computer Science, Speech Technology, Electrical Engineering or a similar discipline.
- Good programming skills in computer languages such as Python or Julia.
- Affinity with machine learning, and specifically deep learning.
- Experience in modern deep-learning frameworks such as PyTorch or TensorFlow.
- Good command of spoken and written English.
- Employment: 32 - 40 hours per week.
- The gross starting salary amounts to €2,395 per month, and will increase to €3,061 in the fourth year (p scale).
- In addition to the salary: an 8% holiday allowance and an 8.3% end-of-year bonus.
- Duration of the contract: you will be appointed for an initial period of 18 months, after which your performance will be evaluated. If the evaluation is positive, the contract will be extended by 2.5 years.
- You will be able to make use of our Dual Career Service: our Dual Career Officer will assist with family-related support, such as child care, and help your partner prepare for the local labour market and with finding an occupation.
- Are you interested in our excellent employment conditions?
For more information about this vacancy, please contact:
Prof.dr. D. van Leeuwen, Full Professor
Tel.: +31 681 888 702