I'm a software engineer. I work on speech recognition at Meta. Before joining Meta, I worked at Speech@Microsoft for a long time. There I have done some cool things:

• The first deep neural networks ever used in production at Microsoft, in our service and on all types of devices that were shipping at the time;

  • Co-invented SVD compression for deep neural networks;

  • Came up with lossless 8-bit quantization on SSE 4.2,  AVX2, ARM Neon (things like asymmetric per-row quantizers you can still find somewhere in the Xbox One code). That was in 2014.

  • I owned the speech recognition Neural Network Runtime for a long time before the ONNX Runtime, CNTK,  Tensorflow, and Pytorch became a thing.

• I briefly worked with Microsoft's not-so-super-secret-anymore new operating system, Midori.  Our first neural networks were shipped on that OS.

• While that was going on, I prototyped the "Xbox, On" wake-on-voice Kinect feature,  and a friend of mine shipped it while I was looking over his shoulder and taught him about speech recognition and WFSTs. The code for "Xbox, On" was the result of my internship project at Microsoft a few years earlier and was based on WFSTs. Oh, I also helped to ship Cortana during that time.

• Later, I mentored the engineering side of the "Hey Cortana" keyword spotter with a different architecture based on DNNs;

• Then I advised the "custom keyword" wake-word detector;

• After that, me and another scientist built and shipped one of the first end-of-speech detection models for snappy query responses. I had always thought this was the way to go, and we proved it with this project.

• Then I pushed out the first end-to-end models that are now replacing WFSTs at high quality, both hardware-accelerated on-device and in the service.

Throughout this time, I also created other innovations which steadily improved our speech recognition engine and service quality in terms of accuracy, latency, and cost. 

When I was attending school in Germany, I worked on Weighted Finite-State Transducers with the late John McDonough. I miss him, and those times we spent together. WFSTs were fun, and 10 years after I joined Microsoft,  they became mainstream there too. Phew!

I've published papers on WFSTs, and hold patents in a few areas related to DNNs, Speech Recognition and Keyword Spotting.

I come from Sofia, Bulgaria - a beautiful city with warm, hospitable, friendly,  and fun people.  I speak Bulgarian, German, and English. I am learning to speak Mandarin.

I like playing volleyball (right-hand opposite hitter and a weak middle) and bullet chess (2100 online, thanks Prateek!). I also have a piano.

Glad you stopped by!