Re:Leon: Open-source, self-hosted personal assistant
by fxtentacle on Sunday September 04, 2022 @06:02PM
(#32718140)
I fully agree and I would love to change that. I mean my company already funded work in that direction... but I sadly predict that we won't have good open source real-time speech recognition anytime soon.
My napkin calculation is that you need about $100k for each attempt at training a Conformer-Transducer. There's a pre-trained NVIDIA model but it appears to have a bad choice of hyperparameters and performance is much worse than what one would expect based on research literature and I believe you're not allowed to execute it on non-NVIDIA hardware.
A skilled team will maybe need 5-10 attempts for discovering a good set of hyperparameters. So the price to create the AI model will likely be around $1 mio. But if you have such large expenses, you have to plan things as a business venture. And that means an open source release is highly unlikely.
(unless, of course, someone like stability.ai is happy to bankroll 200 A100 GPUs for a few months each per target language. In that case, please contact me)
My napkin calculation is that you need about $100k for each attempt at training a Conformer-Transducer. There's a pre-trained NVIDIA model but it appears to have a bad choice of hyperparameters and performance is much worse than what one would expect based on research literature and I believe you're not allowed to execute it on non-NVIDIA hardware.
A skilled team will maybe need 5-10 attempts for discovering a good set of hyperparameters. So the price to create the AI model will likely be around $1 mio. But if you have such large expenses, you have to plan things as a business venture. And that means an open source release is highly unlikely.
(unless, of course, someone like stability.ai is happy to bankroll 200 A100 GPUs for a few months each per target language. In that case, please contact me)