went hunting for open weight english speech models that claim to beat the whisper and parakeet baseline and found more launches than expected: a cohere build, a canary qwen build, a moss model, a korean board topper that is torch only, and a tiny moe released the same week. every one of them claims better accuracy or speed on paper.
tested the four that can run on a cpu, on the same fixed clips as the rest of the ladder. the cohere build transcribes faster than realtime but hallucinated a confident wrong paragraph on a quiet clip. the canary qwen had the cleanest transcripts of anything tested but ran at a fifth of realtime and looped a hallucination on silence. moss made real word errors, request for pull requests. one small model straight up failed, most clips transcribed empty.
also reran the whisper variants and the new parakeet build to keep the ladder honest. whisper turbo is fast but confidently makes up text on unintelligible audio, while the big model outputs silence instead of guessing, which is the smarter failure mode. the new tiny parakeet is accurate but has a quadratic attention cliff, long clips took ten times as long as short ones, and the whole bench shifted under cpu contention from other services on the box.
the verdict is that paper accuracy numbers mean nothing until you run the model on your own files, and none of them toppled the baseline on this machine.