Tyler Weitzman’s Post

Five years ago, my Stanford University grad course project was fine-tuning Google's best text-to-speech model. This week, the model my team built from scratch at Speechify became the #1-ranked TTS model in the world. In 2021 I was a grad student planning to go down the systems track. Then I took Andrew Ng's ML course. The syllabus changed every single week because the field was moving too fast to pin down. That was enough to make me switch paths. That quarter I fine-tuned Tacotron2, Google's state-of-the-art TTS at the time. My TA's feedback is in the screenshot below: "This is fantastic work, Tyler." After the course ended, I decided to keep going. My work wasn't finished. I built the AI lab in Speechify with the goal of building the best text to speech model in the world from scratch. Today, we achieved our goal, 5 years later, with Simba 3.2. It's Speechify's own architecture, designed from the ground up over years of work. And as of this week it's ranked #1 in the world on Artificial Analysis' Speech Arena, ahead of every lab in the field. Google is #2. Yesterday Andrew Ng congratulated me on X. My old TA replied too. People see the leaderboard and think it happened fast. It didn't. It took five years to get here, and it started with a course project. – Written with Speechify Voice Typing