wanted to know which speech to text and text to speech models could actually run on older hardware, the kind with a 1050 ti, so i built a scratch benchmark and ran every candidate against a real corpus with ground truth.
the ladder came out clean. the small parakeet model ran on the old gpu and beat the cpu fallback hard, forty three times realtime instead of twelve. a 38 million parameter model was the fastest but garbled words at nineteen percent error, real garble, not a metric artifact. a 180 million int8 model was slower and worse on accuracy. a streaming 0.6 billion model was unusably slow on cpu. the accuracy winner was the ternary model at about four percent error, which became the fallback.
on the sound side, a seven million parameter model replaced the old fallback and answers through the router in about fifty milliseconds, because it truly synthesizes a short sentence in tens of milliseconds. measured on real clips, both fallbacks now beat or match the primary machine on their own axes. the whole voice path stays under a quarter second whether the big desktop is awake or not.