I ran Whisper on my own box this week. Not the API. The actual open-source mod… · @ReignSwords · TwKit

@ReignSwords

I ran Whisper on my own box this week. Not the API. The actual open-source model, on a 2-core machine I already pay for. 131 words I wrote myself, read aloud as 42 seconds of clean speech. Same clip, two models, plus a 1.5x speed version. My own script was the ground truth, so there's nowhere to hide. tiny (39M params): 2.4 seconds, 3.8% real word errors. base (74M): 3.8 seconds, 3.1% real word errors. 1.5x speed: no accuracy loss at all. That's 17x faster than real time, on a CPU. Cost: $0.00. The API charges $0.006 a minute, and whisper-1 gets switched off in February 2027 anyway. Here's the part the benchmark tables won't tell you. The raw error rate looked like 18-21%. Scary, right? It's a lie. Most of those "errors" were Whisper writing 3,412 instead of "three thousand four hundred and twelve." Correct, just formatted differently. The real mistakes were the same in all four runs. "cache invalidation" came out as "cash and validation" every single time. And "debugging" became "debiting" or "debating." Dev jargon is its blind spot. It has no idea what a cache is. Verdict: for clean, single-speaker audio, run tiny. Base buys you 0.7 points of accuracy for 60% more time. Spend the savings on a better microphone. So who's still paying per minute to transcribe clean audio in 2026 — and what exactly are you buying? 🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥

曝光 53 · 评论 7 · 点赞 0 · 书签 0