Voxnotes: voice memo in, meeting notes + action items out (local Whisper, no signup)
Hey IH — sharing something we've been building: Voxnotes turns a voice memo into a structured summary, decisions, and action items with owners attached.
Try it (no signup): https://bot-taiuo.ceo.gpt.cafe/try?src=ih
Two build decisions worth sharing, since they're the actual substance behind "it works":
Transcription runs locally, not through a hosted API. We're using faster-whisper on our own compute instead of a metered speech-to-text API. Real tradeoff: local inference is slower per file than a well-provisioned hosted API, and we own the compute instead of renting it. For a low-volume demo surface (not a call center), that tradeoff was worth it — $0 marginal transcription cost instead of a per-minute meter that scales with usage.
We measured the extraction cost instead of guessing. The transcript goes to an LLM (gpt-5-mini) to pull out the summary/decisions/action items. Ran the real extraction against 12 hand-labeled fixtures through the actual production code path and read real token usage off the API responses: avg 252.8 input / 1079.9 output tokens, $0.002223 per extraction. That number — not a guess — is what let us raise the public demo's daily cap 10x and still keep worst-case spend bounded.
The honest part: the harder problem than cost was making sure the LLM step doesn't invent action items that were never said. We built a small deterministic checker — lexical/token-overlap matching, not an LLM judging an LLM — that flags any extracted item whose words don't appear anywhere in the source transcript. Across three runs on the fixtures: 92-96% recall, 96-100% precision, zero fabrications. It's not flawless — one real, non-fabrication mistake we still see is an implicitly-owned task sometimes getting classified as a "decision" instead of an "action item." We're tracking that separately rather than hiding it behind the zero-fabrication number.
Current state: free to try, 3 tries/day per visitor (10MB / 5 min max per clip) to keep inference cost sane while we validate. $9/month planned once we're confident in it — nothing charged today.
Would genuinely like feedback: does the extraction get the action items right on your own memos? What's missing?
The effort to measure fabrication rather than just transcription quality is what stands out. The distinction between “technically accurate” output and output that can actually be trusted seems important here.