Georg Heigold

Georg Heigold is a speech and machine-learning researcher at Google who coauthored the 2020 Vision Transformer paper. He earned a physics Diplom at ETH Zurich in 2000, worked as a software engineer at De La Rue in Berne, and wrote a 2010 doctoral thesis on log-linear discriminative modeling for speech recognition at RWTH Aachen University. After joining Google as a Research Scientist in 2010, he coauthored papers on multilingual acoustic models in 2013 and end-to-end speaker verification in 2015. In 2020 he coauthored Slot Attention and the Vision Transformer, and in 2021 he coauthored ViViT, a video transformer. He is a research scientist at Google Research in Munich, and he is the first-listed author of the Massive Sound Embedding Benchmark, presented at NeurIPS 2025.

Now at: Google Research
Badges: Bridge

Timeline

2000–2003

Physics at ETH Zurich

He receives a Diplom in physics from ETH Zurich, Switzerland, in 2000 and works as a software engineer at De La Rue in Berne from 2000 to 2003.
2004–2010

Speech recognition at RWTH Aachen

He is with the Computer Science Department at RWTH Aachen University from 2004 to 2010. In 2009, he coauthors an Interspeech paper with David Rybach, Ralf Schlüter and Hermann Ney on convex optimization with log-linear HMMs for digit string recognition. He writes the doctoral thesis A Log-Linear Discriminative Modeling Framework for Speech Recognition, dated June 2010.
2010

Research scientist at Google

He becomes a Research Scientist at Google in Mountain View, California, in 2010. His research interests include automatic speech recognition, discriminative training, and log-linear modeling.
2013

Multilingual acoustic models

In 2013, he coauthors Multilingual Acoustic Models Using Distributed Deep Neural Networks for ICASSP, with Vincent Vanhoucke and Jeff Dean among the authors. The work trains acoustic models on eleven Romance languages with 10k hours of data in total, and multilingual training gains an average of 7% over monolingual baselines for data-scarce languages and 2% for data-rich ones.
2015

End-to-end speaker verification

In 2015, he coauthors End-to-End Text-Dependent Speaker Verification with Ignacio Moreno, Samy Bengio and Noam Shazeer. The paper formulates verification as a single neural network that maps a test utterance and a few reference utterances to one score, and evaluates it on an internal 'Ok Google' benchmark.
2020

Slot Attention and Vision Transformer

In June 2020, he is a coauthor of Object-Centric Learning with Slot Attention, whose Slot Attention module produces a set of task-dependent abstract representations called slots. In October 2020, he is one of twelve authors of An Image is Worth 16x16 Words, which shows that a pure transformer applied directly to sequences of image patches can classify images very well, and it appears at ICLR in 2021.
2021–2022

ViViT and object-centric video

In 2021, he coauthors ViViT: A Video Vision Transformer with Anurag Arnab, Mostafa Dehghani, Chen Sun, Mario Lučić and Cordelia Schmid, which applies pure-transformer models to video classification and appears at ICCV 2021. In 2022, he coauthors Conditional Object-Centric Learning from Video with Thomas Kipf, Klaus Greff, Alexey Dosovitskiy and others at ICLR.
February 2026

Google Research in Munich

In February 2026, he posts the Massive Sound Embedding Benchmark (MSEB) to arXiv as first of seven authors, after its presentation at NeurIPS 2025. MSEB offers eight core tasks and the new Simple Voice Questions dataset for evaluating the auditory components of multimodal systems. He works at Google Research in Munich, Germany.