Blog · Technology

Speech-to-text is a commodity now. What matters is what happens next

FFmpeg ships Whisper. faster-whisper runs it four times faster on the same hardware. When transcription is free and local, the value moves to accuracy you can check and output a business can use.

10 October 2026 · 2 min read · Soham Mitra

Transcription crossed a line in 2025

In August 2025 FFmpeg 8.0 shipped a Whisper filter, built on whisper.cpp. The changelog gives it one line. As heise put it, it “works locally and does not transfer content to the cloud.” The most widely installed media tool in the world now transcribes speech on the machine it runs on.

Speed is no longer the obstacle

faster-whisper, which Vlocalhost uses, reimplements Whisper on CTranslate2. Its README benchmarks 13 minutes of audio:

So where does the value go?

If transcription is free, local and fast, a product built on it has to be worth something else. We think three things:

1. Accuracy you can check. Speech models and language models both invent things (see our post on hallucinations). Every claim in a note should point at the second it was said.
2. Output a business can use. Nobody's system has a field called “transcript”. It has a demand sheet, a matter log, a consult note. Filling those, in their format, is the work.
3. Where it runs. Local transcription is no longer rare, and we do not claim it is. Open-source projects and commercial products do it too. What differs is what each one does with the words afterwards, and who holds them.

That is why Vlocalhost's focus is VIBE mode, records with citations, rather than the transcript itself. The phone-counter case study shows what that looks like end to end.