Interviews and meetings
Every interview, searchable, on your own disk.
Record or upload. Get a transcript with speakers separated, a note-taking surface tied to the audio timeline, and a library you can actually search years later.
- Who this is for
- Journalists, researchers, clinicians
- Custody
- Self-hosted audio stays local
- Transcription
- Speaker-aware and searchable
- Export
- Markdown, JSON, CSV, SRT, VTT
Who this is for
People whose recordings are covered by an obligation.
Journalists, researchers, and anyone whose recordings are covered by a confidentiality obligation that a cloud transcription service cannot satisfy.
The distinguishing thing is not that you would prefer privacy. It is that uploading the file is a breach, and so the question is not which transcription service to pick but whether you get a transcript at all.
The pipeline
Four steps, and none of them is an upload.
Capture
Record the conversation in the browser on the machine running the instance, or bring a file from a recorder and drop it in. A watch folder does the same thing without you clicking anything.
Transcribe
faster-whisper runs on your hardware with word-level timing on. The local STT path has no vendor per-minute transcription bill, so a three-hour deposition is a scheduling and compute question rather than a usage-meter question.
Separate the speakers
The transcript arrives split by speaker, and you name them once. From then on the names are what search and export use.
The Reader with a two-speaker interview in lanes, one speaker named. Work on it
Press N while the Reader is open to capture a note at the current playhead. Notes attach to a moment rather than to a document, so a note you wrote at 41:20 takes you back to 41:20. Search runs across every recording in the library, not just the one open in front of you.
Notes
A note is a timestamp, not a paragraph in another app.
The language model is off by default. Turning one on is a config change, and every feature that uses it has a defined behaviour for when it is off.
Notes are anchored to the playhead. That is the whole feature and it is the reason the library is usable later: you are not searching for a memory, you are clicking a note and landing on the audio that caused it.
Summaries and reports exist, and they run when you ask for them. On a stock install, with no language model configured, they are extractive — quotes pulled from your own transcript rather than prose written about it. That is worse, and it is also the version that involves no third party at all.
Export
Markdown, JSON, CSV, SRT, VTT, and a report — all per-user.
Everything is exportable per recording, and every export is scoped to the owner by row-level security in Postgres rather than by a filter in the API. A query that forgets whose data it is returns nothing.
The limit worth stating: full-account archive moves are API-led workflows. Export is per object, so a bulk migration means a script against the API rather than one button.
Interview transcription questions
Can interview transcription run without uploading the recording?
Yes. The local STT path runs faster-whisper in your worker against media in your configured storage boundary.
Is transcription quality claimed to beat hosted services?
No. VoxNotch argues for custody and workflow structure. Quality still depends on the recording, model, language, speaker overlap, and review process.
Honest limits
What this does not do for you.
It does not decide what you are allowed to record. Consent, notice, and two-party rules are yours, and nothing in the software checks them.
It does not produce a compliance-grade access log. There is a job record for every unit of work and a migration ledger for every schema change. That is not the same as a record of who read which recording when.
Transcription quality is Whisper’s quality. Accented speech, crosstalk, and bad room audio degrade it exactly as much as they degrade any Whisper deployment.
Start with one interview.
Take a recording you would not have uploaded, and see what comes out.