SMF·NOTTA
Can a prompt replace Notta?
Meeting notes — meeting transcription and AI notes
Exhibit tracking slip
Verdict
Whisper handles the multilingual half well: it detects the language, transcribes it, and will translate to English in the same pass. The gap is everything around the microphone. Notta joins your calls as a bot, records from a phone in your pocket, and syncs the result before you leave the room. A local build transcribes files you hand it, which is a genuinely different workflow even though the model doing the work is comparable.
Exhibit A — The prompt
Received on31.07.2026Build a local multilingual transcription and meeting-notes tool.
Intake: drag in audio or video (WAV, MP3, M4A, MP4, MOV). Normalise everything to 16 kHz mono WAV with ffmpeg before transcription. Support a batch queue, because the realistic use is dropping in a week of recordings at once.
Transcription: run whisper.cpp locally with a user-selectable model size, and expose language handling explicitly rather than hiding it. Three modes: auto-detect, transcribe in a named language, and translate-to-English. Show which mode ran on each transcript, since a mis-detected language produces confident nonsense and the user needs to be able to see why.
Segments: display timestamped, editable segments. Speaker labels start as Speaker 1, Speaker 2 and so on, assigned by a simple energy-and-pause heuristic; the user renames a speaker once and the change propagates through the whole transcript. Do not claim automatic speaker identification — say in the interface that the initial split is a guess.
Notes: on approved transcript text, call an LLM (Anthropic or OpenAI, whichever key is configured) to produce a summary, the decisions reached, open questions, and action items with an owner where one was named. Let the user pick the output language independently of the spoken language, so a French meeting can produce English notes.
Storage: SQLite, holding the recording path, transcript segments, speaker map and generated notes. Full-text search across every transcript. Export Markdown, plain text and WebVTT.
With no API key present, transcription, editing, search and export must all still work; only the generated notes are disabled, with a message saying so.
Out of scope: joining live meetings, any bot that dials into a call, real-time streaming transcription, and mobile capture.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 the meeting bot that joins Zoom, Meet and Teams calls on its own
- B.2 recording from a phone, which is where in-person meetings happen
- B.3 live transcription you can read while the meeting is running
- B.4 automatic speaker identification that is right more often than not
Prior art
Exhibit C — Why people still pay: capture reliability, integrations, and collaboration
Because the transcription is the easy part and the capture is the hard part. A tool that only works on files you remembered to record is a tool you stop using by the third week.
Questions
Can I bring my Notta transcripts over?
Yes. Notta exports transcripts as TXT, DOCX and SRT, and the importer can read SRT directly into timestamped segments. Speaker names come across if they were set in Notta; the original audio does not, so those transcripts become read-only records.
How good is Whisper on languages other than English?
Strong on major European languages and Japanese, noticeably weaker on low-resource languages and on heavily accented speech in any language. The large models are much better than the small ones and much slower. Notta's advantage is not a better model so much as cleaner audio, because it captures the call directly rather than through a room microphone.
What does it cost to run?
Transcription is free after the one-time model download and runs on your own machine. Only the notes step costs anything, and a one-hour transcript summarised through a hosted model comes to a few cents. A year of weekly meetings is a couple of dollars.
What is the one thing that does not survive the rebuild?
Being in the meeting. Notta's bot joins the call whether or not you remembered, and its phone app records the conversation you did not plan to record. This build only ever sees the files you deliberately handed it.
Related tools
Receipt