murmurvault

Local-first recording, transcription and semantic search for meetings and voice notes.

Status: early prototype. The on-disk layout and the CLI may still change.

• Your files, on your disk. Each recording is a directory of plain files: FLAC tracks, JSON transcripts and a meta.json. Folders are real directories. The search index is derived data and can always be rebuilt.
• Microphone and computer audio on separate tracks. Your words are labelled me and the other side others without any model. Optional diarization splits others into individual speakers.
• Live or later. Get a draft transcript while recording, then an accurate pass when you stop. You can also import existing files and transcribe them at any time, with any engine. Every pass is kept.
• Any engine. Built-in faster-whisper, or any OpenAI-compatible /v1/audio/transcriptions server: a hosted API, or a local one such as speaches or whisper.cpp's server.
• Hybrid semantic search. SQLite FTS5 (BM25) keyword search and embedding search, fused with reciprocal rank fusion, then reranked with a cross-encoder. Embedders and rerankers can be local models or API endpoints.
• Folders and tags for organising recordings.
• CLI for scripts and agents (--json everywhere) and a TUI for people. A GUI is planned.

Install

pip install "murmurvault[whisper,search]"          # built-in engine + semantic search
pip install "murmurvault[av]"                       # API engines only; av decodes m4a/mp4/webm imports
pip install "murmurvault[whisper,search,diarize]"  # + speaker diarization (pyannote)
pip install "murmurvault[all]"                      # everything, incl. macOS system-audio capture

System requirements:

• Linux: PortAudio (sudo apt install libportaudio2) for the microphone. For the computer's audio, parec (pulseaudio-utils) or pw-record (PipeWire). Either works on PipeWire systems.
• macOS 14.2+: computer-audio capture uses Core Audio process taps via the macos extra. Grant your terminal both the microphone and the system-audio recording permissions. Alternatively, point audio.system_device at a loopback device such as BlackHole.

Use headphones in calls. Otherwise the far side leaks into your microphone track.

Quick start

murmurvault record --title "Weekly sync" -f work/team -t sync   # Ctrl-C to stop
murmurvault import ~/Downloads/interview.m4a -f research -t interview
murmurvault ls
murmurvault show <id>                 # transcript; --format srt|json|meta
murmurvault search "what did we decide about the budget"
murmurvault tui

Recording consent: tell the people you record. In many places it is a legal requirement.

Commands

IDs can be abbreviated to any unique prefix. With --json, record streams one JSON object per live segment, followed by the final recording.

TUI keys

/ search · r start/stop recording (live draft) · t transcribe · g tags · m move · e rename · F5 refresh · Esc close results · q quit

Configuration

murmurvault config --init writes a commented config.toml to the platform config directory. Set MURMURVAULT_CONFIG to use a different file and MURMURVAULT_VAULT to use a different vault. Example of a local transcription server:

[transcription]
engine = "lan"

[engines.lan]
type = "openai"
base_url = "http://localhost:8000/v1"
api_key_env = ""
model = "Systran/faster-whisper-large-v3"

Diarization uses pyannote/speaker-diarization-community-1. Accept its conditions on Hugging Face, then export HF_TOKEN. Enable it with [diarization] enabled = true, or per run with --diarize.

Vault layout

<vault>/work/team/20261007-101500-a1b2/
  meta.json                   title, tags, tracks, track offsets, active transcript
  mic.flac  system.flac       16 kHz mono tracks
  transcript.live.json        draft made while recording
  transcript.local-small.json accurate pass (one file per engine/model)
<vault>/.murmurvault/index.db search index (derived; `murmurvault reindex`)

Development

uv sync --all-extras --group dev
uv run prek install
uv run pytest

See CONTRIBUTING.md.

License

BSD-3-Clause