Internship - Machine Learning Research - Speech-to-Speech Translation

Speech-to-speech translation is moving from cascaded pipelines to end-to-end models that listen, translate and speak in a single system. Many core questions remain open. How do you translate simultaneously, before the speaker has finished? How do you keep translated speech natural and expressive? How do large multimodal models best combine speech understanding, translation and speech generation? And how should these systems be evaluated?

As an intern, you will own a focused research project on one of these questions, working closely with researchers and engineers on the team. You will have access to large-scale data and compute, and we encourage publishing results at top venues.

Example research directions include:

• Streaming and simultaneous speech translation: policies for when to listen and when to speak, and the trade-off between quality and latency.
• Speech generation for translation: natural, expressive and prosodically appropriate output, including speech tokenizers and audio codecs.
• Multimodal foundation models for speech: how pre-training, fine-tuning and reinforcement learning shape speech-to-speech ability across languages.
• Multilingual scaling: extending to new languages and handling code-switching and unknown source languages.
• Evaluation: automatic metrics and human evaluation methods that reflect real conversational experience.