Mike Seltzer

Also known as Michael L. Seltzer

Mike Seltzer is a speech recognition researcher whose work runs from microphone-array and robust recognition to speech-enabled large language models. He earned a Brown degree in 1996 and a Carnegie Mellon PhD in 2003, then spent years at Microsoft Research, where he coauthored the Computational Network Toolkit report in 2014 and the 2016 human-parity paper on conversational speech. At Facebook AI from 2018, he coauthored Transformer-Transducer and Emformer, and in 2023 he coauthored work that gave LLaMA-7B speech recognition. ISCA named him a Fellow in 2026, and he is Senior Director, AI at Meta.

Now at: Meta
Works: 169
Badges: Bridge

Publication numbers from OpenAlex, as of 22 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

1996–2003

A PhD at Carnegie Mellon

He graduates from Brown University in 1996 after working with Professor Harvey Silverman on the Huge Microphone Array, then works as an applications engineer at Teradyne from 1996 to 1998. At Carnegie Mellon he completes a master's thesis in 2000 and finishes his PhD in July 2003 in the Robust Speech Recognition group, advised by Richard Stern and funded by a Microsoft Research Graduate Research Fellowship. His dissertation designs algorithms that use microphone arrays to improve recognition accuracy in hands-free environments.
2003–2018

Speech researcher at Microsoft

He joins the Speech Technology Group at Microsoft Research in Redmond in 2003. In 2011, he gives an invited talk at ASRU on robust speech recognition, and by July 2017 he is a Principal Researcher presenting a Custom Speech Service demo that tunes acoustic and language models for lecture transcription.
October 2014

The Computational Network Toolkit

In October 2014, he is one of the authors of the Microsoft Research report that introduces computational networks and the Computational Network Toolkit. The framework describes deep neural networks, convolutional and recurrent networks, and LSTMs as directed graphs of computational steps.
2016–2017

Human parity in conversation

In 2016, he is one of eight authors of "Achieving Human Parity in Conversational Speech Recognition", revised in February 2017. Their system reaches error rates of 5.8% on Switchboard and 11.0% on CallHome, edging past professional transcribers at 5.9% and 11.3%.
2018–2020

Streaming speech at Facebook AI

In 2018, he moves from Microsoft to Facebook AI as an Applied Science Manager. In 2019, he coauthors Transformer-Transducer, a streamable end-to-end recognizer with a 6.37% word error rate on LibriSpeech test-clean, and in 2020 he coauthors Emformer, a memory transformer acoustic model for low-latency streaming recognition.
2023

Speech recognition in LLaMA-7B

In 2023, as part of Meta AI, he coauthors "Prompting Large Language Models with Speech Recognition Abilities". A small audio encoder attached to the open-source LLaMA-7B lets the model perform speech recognition, and on Multilingual LibriSpeech it outperforms monolingual baselines by 18%.
2026

An ISCA Fellowship

The International Speech Communication Association names him a 2026 Fellow, with a citation for contributions to environmental robustness of speech recognition systems and for sustained leadership of production speech solutions in industry.
As of October 2026

Senior Director of AI at Meta

In 2023, he becomes Senior Director, AI at Meta. In February 2026, he is among the coauthors, from Meta Superintelligence Labs and the Chinese University of Hong Kong, Shenzhen, of "Scaling Speech Tokenizers with Diffusion Autoencoders", which introduces SiTok, a diffusion autoencoder that turns speech into tokens for speech language models.