I wanted to return to a specific idea in a recording and know who said it. I designed and developed Bardo to bring audio, transcripts, and speakers together in a native app—first for Mac, then for Windows. Transcription runs entirely on-device, without uploading recordings to a server.
The conversations, names, and company shown in the screenshots are fictional.
Getting started
Choosehowtotranscribe.
I structured first-time setup around three options: Fast, Balanced, and Accurate. Each explains its requirements and how audio is processed on Mac.
Transcription
Followlivetranscription.
I kept progress visible in both the conversation and the library. Words appear as they are recognized.
Listening
Seewho'sspeaking.
I separated Mac system audio from the user's microphone into two tracks, with speaker colors and controls for listening to each.
Bookmarks
Saveamomentwithitscontext.
I added moments and sections to both the waveform and transcript, with notes attached to specific points.
Search
Searchacrossyourlibrary.
I connected search to titles, transcripts, and participants. Highlighted matches help locate each result within a recording.
Speakers
Identifytheparticipants.
I added voice samples so speakers can be recognized, named, or merged if the same person was detected more than once.
Windows
Thesameapp,builtforWindows11.
I designed the Windows version with WinUI 3 and Fluent instead of imitating macOS. Both share a library format, so a recording copied from Mac opens on PC.
Record on PC
YourvoiceandyourPC'saudio,eachinitsowncolor.
I captured microphone and system audio separately, with sections and moments that can be marked while recording. Transcription happens locally on the PC.
Implementation
HowIbuiltit.
Preserve the recording
I stored audio in recoverable chunks and added a capture log. I tested interruptions because preserving the recording is a prerequisite for transcription.
A native Mac app
I built the interface with SwiftUI and AppKit, integrating WhisperKit and SpeakerKit for transcription and speaker diarization. I brought transcripts, audio, search, and bookmarks together in the player.
A native Windows app
I built the Windows version in C# with WinUI 3 and Fluent, without sharing Swift code. It uses Whisper.net with the same models and sherpa-onnx for speaker detection, and stores the library in the same format as the Mac app.
macOS 15 or later · Apple Silicon · Windows 11 (x64 and ARM)
Runs on Apple Silicon Macs with macOS 15 or later, and on Windows 11 x64 or ARM PCs. Models must be downloaded on first use. Apple's recognition features require macOS 26. Archiving and the recording indicator are currently Mac-only. Bardo transcribes conversations; it does not generate meeting minutes with a language model.
