Introduction
Your phone or browser sends audio to the DictaWhisper server you configure. That computer transcribes it. If refinement is enabled, transcript text goes to the configured Ollama host.
DictaWhisper is a local web app for a searchable voice journal. Record in the browser, import audio, or sync a phone folder. Read and search the transcript beside its recording, replay passages, and review possible actions in MayDos. Transcription uses Bark Pro (English, default) or faster-whisper. Optional cleanup uses ollanet on the host you configure.
Open the transcript as text. Use the accompanying JSON when you need the structured recording and transcription data.
The name is dicta (dictation, a dictaphone) plus Whisper. Audio stays on this computer. The .json next to each recording is the note: no database, no account.
Starting with 0.1.11, the npm package ships compiled JavaScript and the inbox UI. Earlier releases through 0.1.10 do not start from npm installations. The selected transcription engine and its Python runtime remain separate prerequisites. There is no installer. Hardware and runtimes belong on Install before the clone.
Why it exists
Voice notes are easy to make and hard to keep. Phone recordings pile up as undated blobs. Cloud speech-to-text wants the audio. Desktop Whisper dumps a wall of filler words and repeated phrases. A journal you can reread needs three things:
- Capture that fits your tools: record or drop a file in the browser. Syncthing from a phone is optional.
- Local transcription: Bark Pro for English or Whisper, with word timing estimates used to connect the transcript to playback. Check important details against the audio.
- Usable notes: cleaned prose, tags, playback that follows the cleaned paragraphs, files you can copy and back up.
What it is good at
Files are the database. Each note is audio + audio.json. You can rsync the tree, open a sidecar in an editor, or point pnpm retranscribe at one month. The file path is the note's identity, so there is nothing to export.
Choose the transcription engine. Bark Pro is the English default and loads per recording. Whisper keeps its worker loaded between jobs. Set the model in Settings → Transcription after completing the engine setup.
Audio is transcribed on this computer. The selected transcription engine runs here. If you point ollanet.machine at Ollama on another machine on your own network, only cleanup text goes there. If ollanet is unreachable, doctor and /health report it and you still have the raw words.
Playback aligned to word timestamps. Cleanup drops fillers and collapses repeated phrases, but playback cues map back to the transcription engine’s word timing estimates rather than being inferred from the cleaned wording.
Custom vocabulary and stable tags. whisper.promptTerms go into Whisper's initial prompt. Tags already in your inventory are fed into cleanup so the model does not spell them a new way every week.
Privacy and network boundary
- No telemetry. Nothing contacts an external analytics or telemetry service.
- Audio and transcript files stay here. The selected local engine processes audio, and the notes stay next to the audio.
- Cleanup text goes only to
ollanet.machine. Cleanup sends transcript text to that one host and is skipped when it is empty. - Search embeds on this computer by default. Search indexing sends note text (up to 8,000 characters per note) to an Ollama embedding model on this computer. It uses another host only if you set
journal.embedHost:machinefor theollanet.machinehost, a host name for that host, oranyfor the first host ollanet discovers. If no allowed host has an embedding model, nothing is embedded, search matches words only, and doctor and the inbox Tools panel say why. Withjournal.searchset tolex, nothing is embedded at all.journal.embedModelnames the model, not the host. - Loopback by default. API (
8008) and UI (7777) bind to127.0.0.1. Withhttp.tailscale, the inbox also binds to this machine's Tailscale address, not to every interface. - Path allowlisting. File routes resolve realpaths and reject anything outside
watch.rootsandbrowserDropFolderwith a403.
By default the API binds to 127.0.0.1. Turn on http.tailscale and the inbox also listens on this machine's Tailscale address (100.x, plus MagicDNS when you have it) so a phone on the same tailnet can open it. The API stays on loopback; Vite proxies to it.
Files on disk
| Item | Default | What it is |
|---|---|---|
| Browser drop folder | ./data/audio-files |
Recordings and drops from the inbox |
| Watch roots | none | Optional phone or dump folders |
| Notes file | <audio>.json next to the audio |
Raw text, segments, cleaned text, tags |
| Journal index | ./data/journal.sqlite |
Search index, not the source of truth |
Copy the audio and its JSON together and you copied the note.
Next
- Install — hardware, clone, doctor
- Quick start — record, drop, recovery
- Config — env and
config.json