Applied Scientist - Internship

Microsoft Teams is the hub for teamwork that integrates all the people, content, and tools your team needs to be more engaged and effective. It is core to Microsoft’s modern work, modern life & modern education value prop. We are reinventing the way people communicate and work together across the globe. We are looking to hire a PhD (or published MSc) candidate for a 12-week internship to join CMD Labs – an applied science team within Microsoft Teams – to work on improving transcription accuracy by applying existing or novel research and leveraging training, fine tuning, and prompt engineering of speech transformer models, as well as LLMs and audio-enabled foundations models as post-processing and re-scoring modules. Our flagship AI applications for Teams Meetings such Meeting Agent, Meeting Copilot and Intelligent Recap are all fully dependent on an accurate meeting transcription as the primary grounding data. When used as grounding data for AI, transcription quality can significantly affect AI reliability. For instance, the importance of named entities - names of people, projects, products, companies and places - is often the most important, and yet the most challenging for the transcription engine since the names might not be a part of the model's training data. An important part of the challenge is to unravel the aspects of accuracy that affect AI reliability the most, and thu setting relevant metrics and objectives beyond WER. The intern will be onboarded to our evaluation pipeline code processing real internally donated meetings and work on improving existing algorithms as well as proposing novel solutions to the problem based on recent academic literature. The work done in the internship will contribute towards the algorithm that the engineering team will implement in production. Given substantial scientific novelty of the approach and results, collaboration on a mutual publication is encouraged.