Dylan Fox

Fox studied at George Washington University, where a 2013 university magazine profile described his business administration and public-policy studies. It also covered Crowdvance, an online fundraising service he developed with partners. His professional profile lists experiments including ClaimYourName, which automated setting up a personal website on a custom domain, and BitJournal, an encrypted journal.[5][6]

Making speech models accessible to developers

Fox says he started AssemblyAI in 2017. A Bluejay interview describes his natural-language-processing work at Cisco. AssemblyAI offers developers speech models through infrastructure they can incorporate into applications.[4][7][8]

Conversations with AI founders

Fox hosts Assembly Required, a company interview series featuring builders such as Ironclad’s Jason Boehmig, Synthesia’s Victor Riparbelli and Pinecone’s Edo Liberty. The discussions cover product strategy, customer trust and the timing of technical opportunities. His contribution is as interviewer and fellow founder. The series complements his writing about speech evaluation by examining the decisions that turn model capabilities into products people will use.[2]

Why transcription accuracy needs context

In Beyond WER, Fox argues that word error rate is an incomplete guide to speech-to-text quality. The metric counts differences from a reference transcript, treating a harmless formatting difference and a meaning-changing mistake as comparable errors. He also presents cases where reference transcripts omit words, contain incorrect names or handle multiple languages inconsistently. His argument is that developers should inspect errors in relation to the application using the transcript. This is a company founder’s analysis of benchmark examples, not an independent finding that one provider is universally better.[3]

Listening inside a working application

Fox’s Bluejay interview examines failures that arise when speech recognition becomes part of a voice agent. One example involves background speech being treated as the main speaker. The conversation frames speech-to-text as a listening component in a larger interaction, where deciding whose words matter can be as important as spelling those words correctly. Fox argues for starting from the user’s experience when evaluating competing voice architectures. The practical question is whether the whole interaction works under realistic conditions.[4]