Speech and language modeling (Gemini), dialogue agents, long-form speech synthesis, streaming models, streaming infrastructure.

I lead a small team (7 people, including 20%ers) building neural networks for end-to-end speech synthesis using techniques from generative modeling and probabilistic machine learning.I'm the co-inventor of Tacotron 🌮🤖:, the first "end-to-end" text-to-speech model (though, it depends on how you define your ends ;)) and made foundational contributions to expressive prosody modeling in an end-to-end setting by treating prosody as a latent variable to be inferred during training.Check out our publications for more details: https://google.github.io/tacotron/For an overview of our 2018 work on global prosody/style modeling, check out this blog post:https://ai.googleblog.com/2018/03/expressive-speech-synthesis-with.htmlI've contributed to the launch of Tacotron (and similar models) in various Google products.In my 20% time, I work on TensorFlow. My goal is to make TensorFlow easy to use for signal processing and audio use cases.* I authored tf.signal, a library for signal processing that is GPU capable and has gradient support (including GPU/TPU accelerated FFT/STFT support with gradients).* I added Audio summaries to TensorFlow and TensorBoard to make it easy to listen to audio generated by your models.* I improved complex number GPU support across many core ops (e.g. matmul) to better support complex-valued neural networks.Check out my commit history for more: https://github.com/tensorflow/tensorflow/commits/master?author=rryanOther projects I've contributed to while on the Sound Understanding team:* Audio event classification with CNNs.* Speaker identification.* Adversary-resistant audio fingerprinting / similarity (distance metric learning) e.g. for YouTube Content ID.

Before transferring to Mountain View and joining Machine Perception, I spent 4.5 years in Search Quality working to improve the quality of your web searches via ranking improvements and search features.My main focus during this time was the Knowledge Graph, which I was involved with the launch of in 2012.http://insidesearch.blogspot.com/2012/05/introducing-knowledge-graph-things-not.htmlhttps://googlewebmastercentral.blogspot.com/2013/05/using-schemaorg-markup-for-organization.htmlI worked on a project at the intersection of image understanding and the knowledge graph. Every image you see in Knowledge Graph features on Search is powered by this project. From 2013 to 2015 I led this team and helped spread our work around Google with integrations in Google Now, Google Play, Ads, Trends, Inbox/Calendar, Travel/Flights, Maps/Geo, Research, and more.In 2013 and 2014, I collaborated closely with the Google Inbox and Calendar teams to bring Knowledge Graph smarts to their product before the launch of Inbox in October 2014.https://googleblog.blogspot.com/2014/10/an-inbox-that-works-for-you.htmlKnowledge Graph integration features prominently when you receive flight/hotel/restaurant/car reservation emails, make reminders, and so on.My first project at Google was ranking medical conditions based on search queries to suggest possible problems for queries like [why does my back ache] or [chest pain].http://insidesearch.blogspot.com/2012/02/improving-health-searches-because-your.html

Mixxx is the world's most popular open-source DJ software project. It is downloaded by millions yearly and used by both professional and aspiring DJs alike.

Co-founded an organization (now called Sana) that makes open-source software to deliver medical care in developing nations via mobile phones. Moca is currently working with industry partners such as GE, Partners in Health, McKesson, and the MIT Media Lab to deliver medical solutions around the world.

Feature development on Google Maps

I was an intern in the DirectX Driver group.

Summer internship. I helped write a program to algorithmically judge the quality of a voice recording from a VoIP call.