Damwha
Knows who is speaking
Voiceprint-based speaker identification. Enroll a speaker once and they're recognized in every later conversation.
Free and open source · macOS
Wait, who said that?
Record a conversation and Damwha keeps track of who said what, and when. Search for it and jump straight to that moment in the original audio.
v0.4.1 · 490 MB app · Apple Silicon (M1+) · macOS 15+Plus 8–40 GB of models, downloaded on first useLive demo: read-only, no sign-up · sample conversations are in Korean
Conversations pile up. Finding what was said doesn't get easier.
Memory fades, and finding one line in a long recording is hard. Damwha tells speakers apart by their voices, and once it knows someone, it recognizes them in every other conversation too. Every line is saved with its text, audio and the turns around it.
Utterance jump
Search every conversation, press Enter, and land on the exact second — the line highlighted, the audio cued up.
What else it does
⌘K search, everywhere
Hybrid search — meaning and keywords — over every utterance and conversation, from any screen.
Lenses: decisions, action items, promises
A local LLM pulls them out of each conversation, each with a link back to the evidence. Edit freely — re-extraction keeps what you touched.
Tuned to your Mac
Damwha reads your chip and memory and recommends a preset. Summaries come in the transcript's language, Korean or English — a Korean meeting can be summarized in English.
Private by design
No cloud ML. Transcription, speaker identification, search and summaries all run on your Mac, and voiceprints stay on disk. Models download once, the first time they're needed — your recordings never leave the machine.
- Audio
- Normalize
- Voice activity
- Diarization
- Speaker ID
- Whisper
- Search index
- Lenses ∥ summary
Requirements
- Mac
- Apple Silicon (M1 or later), macOS 15 or later
- Disk for models
- About 8 GB (Light) · 20 GB (Standard) · 40 GB (Quality), downloaded on first use
- Hugging Face token
- Needed only for speaker diarization, with the model terms accepted on Hugging Face. Add it inside the app — Damwha opens without one.
- Display language
- Korean and English. English UI is rolling out — some screens are still in Korean.
- Conversation language
- Transcription: auto, Korean, English, Japanese, Chinese. Summaries: the transcript's language, Korean or English.
Download for Mac
v0.4.1 · 490 MB app · Apple Silicon (M1+) · macOS 15+Plus 8–40 GB of models, downloaded on first use
Before you record
Damwha shows no recording notice. It stores a voiceprint for everyone in the conversation, not just you. Getting consent — and following the law where you are — is on whoever runs it.
Questions
Is it free?
Yes. Damwha is open source under the MIT license.
Do I need an internet connection?
Only to download models the first time they're used and to check for updates. Recording, transcription and search work offline.
Does it run on Intel Macs or Windows?
No. The ML pipeline runs on Apple's MLX, which needs Apple Silicon.
Is the app in English?
English UI is rolling out. Settings and the app menu are translated, and the remaining screens follow in upcoming releases. Conversations in Korean, English, Japanese and Chinese can be transcribed, and summaries can be written in Korean or English.
Can it record live?
Yes. Live recording shows a running transcript preview, and the full pipeline runs when you stop.
How do updates work?
The app checks GitHub Releases and tells you when a new version is out. Download the new DMG and replace the app — your data stays.
Can I undo an update?
Yes. Before a new version upgrades the database, Damwha keeps a snapshot, and the app menu can restore it.
Are the demo voices real people?
No. The sample conversations are AI-generated (Google NotebookLM Audio Overviews), processed by the real pipeline.