ownscribe

Local-first meeting transcription and summarization CLI. Record, transcribe, and summarize meetings and system audio entirely on your machine – no cloud, no bots, no data leaving your device.

System audio capture requires macOS 14.2 or later. Other platforms can use the sounddevice backend with an external audio source.

Table of Contents

• Privacy
• Features
• Requirements
• Installation
• Usage
• Configuration
• Summarization Templates
• Speaker Diarization
• Acknowledgments
• Contributing
• License

Privacy

ownscribe does not:

• send audio to external servers
• upload transcripts
• require cloud APIs
• store data outside your machine

All audio, transcripts, and summaries remain local.

Features

• System audio capture — records all system audio natively via Core Audio Taps (macOS 14.2+), no virtual audio drivers needed
• Microphone capture — optionally record system + mic audio simultaneously with --mic
• WhisperX transcription — fast, accurate speech-to-text with word-level timestamps
• Speaker diarization — optional speaker identification via pyannote (requires HuggingFace token)
• Pipeline progress — live checklist showing transcription, diarization sub-steps, and summarization progress
• Local LLM summarization — structured meeting notes with a built-in model (Phi-4-mini); also supports Ollama, LM Studio, or any OpenAI-compatible server
• Summarization templates — built-in presets for meetings, lectures, and quick briefs; define your own in config
• Ask your meetings — ask natural-language questions across all your meeting notes; uses a two-stage LLM pipeline with keyword fallback
  
• Silence auto-stop — automatically stops recording after sustained silence (default: 5 minutes, configurable)
• One command — just run ownscribe, press Ctrl+C when done, get transcript + summary

Requirements

• macOS 14.2+ (for system audio capture)
• Python 3.12+
• uv
• ffmpeg — brew install ffmpeg
• Xcode Command Line Tools (xcode-select --install)

Summarization works out of the box — a local model (Phi-4-mini, ~2.4 GB) downloads automatically on first run. Optionally, you can use Ollama, LM Studio, or any OpenAI-compatible server instead (see Configuration).

Works with any app that outputs audio through Core Audio (Zoom, Teams, Meet, etc.).

Tip: Your terminal app (Terminal, iTerm2, VS Code, etc.) needs Screen Recording permission to capture system audio. Open the settings panel directly with:

open "x-apple.systempreferences:com.apple.preference.security?Privacy_ScreenCapture"

Enable your terminal app, then restart it.

Installation

Quick start with uvx

uvx ownscribe

On macOS, the Swift audio capture helper is downloaded automatically on first run.

From source

# Clone the repo
git clone https://github.com/paberr/ownscribe.git
cd ownscribe

# Build the Swift audio capture helper (optional — auto-downloads if skipped)
bash swift/build.sh

# Install with transcription support
uv sync --extra transcription

Usage

Record, transcribe, and summarize a meeting

ownscribe                    # records system audio, Ctrl+C to stop

This will:

1. Capture system audio until you press Ctrl+C (or auto-stop after 5 minutes of silence)
2. Transcribe with WhisperX
3. Summarize with your local LLM
4. Save everything to ~/ownscribe/YYYY-MM-DD_HHMMSS/

On first run, WhisperX / pyannote and the summarization model may download model files. ownscribe shows a Preparing models step and best-effort download progress in the TUI while this happens. Use ownscribe warmup to pre-download all models.

Options

ownscribe --mic                               # capture system audio + default mic (press 'm' to mute/unmute)
ownscribe --mic-device "MacBook Pro Microphone" # capture system audio + specific mic
ownscribe --device "MacBook Pro Microphone"   # use mic instead of system audio
ownscribe --no-summarize                      # skip LLM summarization
ownscribe --diarize                           # enable speaker identification
ownscribe --language en                       # set transcription language (default: auto-detect)
ownscribe --model large-v3                    # use a larger Whisper model
ownscribe --format json                       # output as JSON instead of markdown
ownscribe --no-keep-recording                 # auto-delete WAV files after transcription
ownscribe --template lecture                  # use the lecture summarization template
ownscribe --silence-timeout 600               # auto-stop after 10 minutes of silence
ownscribe --silence-timeout 0                 # disable silence auto-stop

Subcommands

ownscribe devices                  # list audio devices (uses native CoreAudio when available)
ownscribe apps                     # list running apps with PIDs for use with --pid
ownscribe warmup                   # prefetch WhisperX/pyannote models before a meeting
ownscribe transcribe recording.wav # transcribe an audio file (saves alongside the input)
ownscribe summarize transcript.md  # summarize a transcript (saves alongside the input)
ownscribe resume ./2026-02-20_1736 # resume a failed/partial pipeline in a directory
ownscribe ask "question"           # search your meetings with a natural-language question
ownscribe config                   # open config file in $EDITOR
ownscribe cleanup                  # remove ownscribe data from disk

Use warmup ahead of time to avoid first-run model download delays while recording:

ownscribe warmup                    # prefetch Whisper model (+ diarization if enabled in config)
ownscribe warmup --language en      # also prefetch alignment model for English
ownscribe warmup --with-diarization # force diarization warmup for this run

Searching Meeting Notes

Use ask to search across all your meeting notes with natural-language questions:

ownscribe ask "What did Anna say about the deadline?"
ownscribe ask "budget decisions" --since 2026-01-01
ownscribe ask "action items from last week" --limit 5

This runs a two-stage pipeline:

1. Find — sends meeting summaries to the LLM to identify which meetings are relevant
2. Answer — sends the full transcripts of relevant meetings to the LLM to produce an answer with quotes

If the LLM finds no relevant meetings, a keyword fallback searches summaries and transcripts directly.

Configuration

Config is stored at ~/.config/ownscribe/config.toml. Run ownscribe config to create and edit it.

[audio]
backend = "coreaudio"     # "coreaudio" or "sounddevice"
device = ""               # empty = system audio
mic = false               # also capture microphone input
mic_device = ""           # specific mic device name (empty = default)
silence_timeout = 300     # seconds of silence before auto-stop; 0 = disabled

[transcription]
model = "base"            # tiny, base, small, medium, large-v3
language = ""             # empty = auto-detect

[diarization]
enabled = false
hf_token = ""             # HuggingFace token for pyannote
telemetry = false         # allow HuggingFace Hub + pyannote metrics telemetry
device = "auto"           # "auto" (mps if available), "mps", or "cpu"

[summarization]
enabled = true
backend = "local"         # "local" (built-in, no server needed), "ollama", or "openai"
model = "phi-4-mini"      # local: "phi-4-mini", path to GGUF, or hf:owner/repo/file.gguf; ollama/openai: model name
# host = "http://localhost:11434"  # only for ollama/openai backends
# template = "meeting"    # "meeting", "lecture", "brief", or a custom name
# context_size = 0        # 0 = auto-detect from model; set manually for OpenAI-compatible backends

# Custom templates (optional):
# [templates.my-standup]
# system_prompt = "You summarize daily standups."
# prompt = "List each person's update:\n{transcript}"

[output]
dir = "~/ownscribe"
format = "markdown"       # "markdown" or "json"
keep_recording = true     # false = auto-delete WAV after transcription

Precedence: CLI flags > environment variables (HF_TOKEN, OLLAMA_HOST) > config file > defaults.

Summarization Templates

Built-in templates control how transcripts are summarized:

Template	Best for	Output style
meeting	Meetings, standups, 1:1s	Summary, Key Points, Action Items, Decisions
lecture	Lectures, seminars, talks	Summary, Key Concepts, Key Takeaways
brief	Quick overviews	3-5 bullet points

Use --template on the CLI or set template in [summarization] config. Default is meeting.

Define custom templates in config:

[templates.my-standup]
system_prompt = "You summarize daily standups."
prompt = "List each person's update:\n{transcript}"

Then use with --template my-standup or template = "my-standup" in config.

Speaker Diarization

Speaker identification requires a HuggingFace token with access to the pyannote models:

1. Accept the terms for both models on HuggingFace:
   • pyannote/speaker-diarization-3.1
   • pyannote/segmentation-3.0
2. Create a token at https://huggingface.co/settings/tokens
3. Set HF_TOKEN env var or add hf_token to config
4. Run with --diarize

On Apple Silicon Macs, diarization automatically uses the Metal Performance Shaders (MPS) GPU backend for ~10x faster processing. Set device = "cpu" in the [diarization] config section to disable this.

Acknowledgments

ownscribe builds on some excellent open-source projects:

• WhisperX — fast speech recognition with word-level timestamps and speaker diarization
• faster-whisper — CTranslate2-based Whisper inference
• pyannote.audio — speaker diarization
• llama.cpp / llama-cpp-python — local LLM inference
• Ollama — local LLM serving
• Click — CLI framework

Contributing

See CONTRIBUTING.md for development setup, tests, and open contribution areas.

License

MIT