Get started
Margin requires an Apple Silicon Mac (M1 or later) running macOS 14 or newer. Intel Macs are not supported. Choose a recording or drag it into the app. Whisper large-v3-turbo is the default model. Turn Identify speakers on when you want voice labels, then choose Transcribe recording.
On first use, Margin downloads the speech models you selected from Hugging Face. Keep the internet connection available and allow several GB of free storage. The progress card shows the download stage and received bytes. Once the models are cached, transcription works offline.
Review the transcript and edit its topic, context, date or speaker names. Under export options, the clean and timestamped Markdown notes both start selected. Each has its own checkbox, so you can export either one or both. Open More formats to add plain text (TXT), SRT subtitles or JSON. These start turned off. Choose a destination folder, then select Export notes. Margin writes only the formats you selected and remembers your choices for next time.
The two Markdown files link to each other only when both are exported. If you turned Identify speakers off for the recording, the reader and Markdown, TXT and SRT exports omit speaker labels. Paragraphs and optional timestamps remain.
Download appears slow
Check the current stage and received bytes in the progress card. Downloading and preparing models can take longer than transcription. If the connection drops, Margin retries automatically and reuses downloaded data where possible. If it still cannot connect, restore your connection and choose Retry download. Completed files are checked before use.
Cancel stops the current task without removing your recording or cached model files. If a model file is unavailable or storage is full, address the displayed issue before trying again. The first use of a different model needs that model’s files too.
Recording cannot be opened
Choose the file again through Margin’s file picker, especially if it moved or its access permission changed. Make sure the recording is available locally rather than only as a cloud placeholder. The file must include audio that macOS can decode. Encrypted or protected media is not supported. For an unsupported recording, export a WAV or M4A copy in the app that created it.
Recording is too long for this Mac
Margin checks a conservative audio-memory budget based on your Mac and selected model. The maximum is four hours, but a lower-memory Mac can have a shorter limit. If Margin asks you to split the recording, export shorter parts in the app that created it and transcribe them separately. Other apps also use memory. Close memory-heavy apps if a recording within the limit still cannot finish.
Export folder needs permission
Choose Change… beside Save to, then select your Obsidian vault or another writable folder. Margin stores that permission for future launches. If the folder moved, was disconnected or no longer allows access, selecting it again restores permission. Every export uses a new folder so earlier notes remain intact.
Earlier transcripts are missing after upgrading
A sandboxed app has separate local storage. Open How Margin works → Import earlier Margin library… and choose the earlier version’s jobs folder. Margin copies saved transcript results and leaves the earlier library in place. This does not transfer model downloads or move your original recordings.
Find your local files
Open How Margin works → Show local data… to reveal Margin’s local storage. It contains job records, results, diagnostic logs, model files and possibly decoded working audio. These files can contain private content. Export anything you want to keep before managing app data in Finder.
Report a bug
To report a bug, email pocketrenaissancestudio@gmail.com with the app version, your macOS version and a short description of what happened. You can find the app version in How Margin works. Please do not send a recording or transcript unless you intend to share it. Review a diagnostic log before attaching it because it can include file paths and job details.
Speech recognition and speaker labels can be wrong. Review important information against the original recording before sharing or relying on it.