Mishearings: Hacking a tiny ASR model to write Dadaist poetry
What happens when you force a tiny speech recognition model to find patterns in mangled speech?
I'm Evan, a PhD AI Research Engineer. I build on-device AI/ML systems that leverage audio, speech, and language models to solve human-centered problems. In my free time I'm a musician, creative coder, cyclist, and photographer. This site hosts a mix of my personal and professional projects. All views expressed here are my own.
Featured Projects
- Feb '25 – Mishearings: Hacking a tiny ASR model to write Dadaist poetry
- Jan '25 – A private ambient listener for homes and offices
- Oct '24 – Moonshine: Industry-leading speech recognition models for edge devices
- May '24 – Teaching things to think: Bootstrapping on-device reasoning for smart devices
- May '23 – Sasha: Introducing LLMs for reasoning in smart homes
- Feb '18 – Turning sea levels into music with CSV to MIDI
All Posts
Moonshine: Industry-leading speech recognition models for edge devices
I'm a core contributor to these speech-to-text models released by Useful Sensors. They're 5x faster than Whisper with higher accuracy!
Teaching things to think: Bootstrapping on-device reasoning for smart devices
How can you train smart devices to reason about their actions and state? That's what I investigated in this IEEE PerCom '25 paper.
Adding eurorack features to a tape recorder
A brief write-up of my hardware hacking project to add independent audio outputs and gates to a TASCAM Porta 03 tape recorder.
Sasha: Introducing LLMs for reasoning in smart homes
What happens when you put LLMs in control of smart homes? I investigated this in my ACM IMWUT/Ubicomp '24 paper.
Using LLMs to control a smart home
Experimenting with GPT models to build a smarter smart home assistant.
