Navdeep Jaitly

Timeline

1. By 2008

   Computational biology and proteomics

   After a double major in mathematics and computer science at Hanover College, he starts a master's at the University of Waterloo and moves into its computational biology group in 1999. He then spends about eight years as a computational biology researcher at a biotech startup and at the Pacific Northwest National Laboratory, building statistical methods for high-throughput proteomics.

2. 2008–2011

   A PhD under Geoffrey Hinton

   He joins the University of Toronto in 2008 for a PhD in computational biology, then moves to the machine learning group, where Geoffrey Hinton agrees to supervise him. His 2011 ICASSP paper on restricted Boltzmann machines shows that features learned from raw speech waveforms outperform standard MFCC features, and his thesis is titled Exploring Deep Learning Methods for Discovering Features in Speech Signals.

3. 2011

   Neural networks in Voice Search

   During a 2011 internship at Google, he plugs deep neural network acoustic models into Google's speech recognizer and cuts the word error rate on a small dataset from 23% to 18%. A version of the model goes into production for transcribing Google Voice Search queries.

4. 2015

   Listen, Attend and Spell

   As a research scientist on Google's Brain team, working on deep learning models for sequences, he coauthors Listen, Attend and Spell in 2015 with William Chan, Quoc V. Le and Oriol Vinyals. The network transcribes speech directly to characters and reaches a 14.1% word error rate on a subset of the Google voice search task without a dictionary or language model.

5. 2015

   Pointer Networks and adversarial autoencoders

   In 2015, he coauthors Pointer Networks with Oriol Vinyals and Meire Fortunato, an architecture that uses attention as a pointer to select a member of the input sequence as the output. He also coauthors Adversarial Autoencoders with Alireza Makhzani, Jonathon Shlens, Ian Goodfellow and Brendan Frey, which use a generative adversarial network to match an autoencoder's code to a chosen prior.

6. 2017

   Tacotron 2 speech synthesis

   In 2017, he is one of the authors of Tacotron 2, with Jonathan Shen, Ruoming Pang, Ron J. Weiss and colleagues. The system maps characters to mel spectrograms with a sequence-to-sequence network and synthesizes waveforms with a modified WaveNet, reaching a mean opinion score of 4.53 comparable to 4.58 for professionally recorded speech.

7. 2023

   Matryoshka diffusion models at Apple

   At Apple Machine Learning Research, he coauthors Continuous Pseudo-Labeling from the Start, a speech recognition method published in March 2023 that generates pseudo-labels from the first step of training, with Tatiana Likhomanenko and others. In October 2023, he coauthors Matryoshka Diffusion Models with Jiatao Gu, Shuangfei Zhai, Yizhe Zhang and Josh Susskind, an end-to-end framework for high-resolution image and video synthesis.

8. 2025–2026

   SimpleFold and SpeakStream at Apple

   In April 2025, he is a research scientist at Apple Machine Learning Research and gives a keynote at Purdue's Applied AI Day on the drawbacks of data tokenization in generative AI. In September 2025, he coauthors SimpleFold with Yuyang Wang, Jiarui Lu, Josh Susskind and Miguel Angel Bautista, the first flow-matching protein folding model built only from general-purpose transformer layers, and Apple presents it at ICLR in Rio de Janeiro in April 2026. In July 2026, he is a coauthor of SpeakStream, a streaming text-to-speech paper shown at the ICML 2026 audio workshop in Seoul.