Steno 0.7.0

Steno v0.7.0

Speakers

• Individual speaker labels (macOS) — Steno separates the voices sharing one side of a call, on-device, once the recording stops. A meeting with several remote participants now reads as Speaker 2, Speaker 3 instead of collapsing everyone into a single [Others]. Works whether people are gathered around one microphone or joining remotely.
• Speaker identification — Name a speaker once and Steno suggests that person in later meetings. Off by default. It creates a numerical voice profile, so turn it on under Settings → AI → Speaker identification only if you have told the people you record and are authorised to do so. Anonymous per-meeting splitting keeps working with it off.
• People management — Settings → People lists everyone you have named, plays the voice sample behind each profile, and deletes a person and their profile permanently. Recordings and transcripts are stored separately and are not deleted with a profile.
• Review who is who — When Steno is unsure, a review panel in the note offers excerpts you can play. Confirm or rename a speaker, mark a cluster as more than one person, or keep the generic label. It tells you what a delete costs before you confirm it.

Notes and transcripts

• Readable transcript export — Exported transcripts open with a clean conversation view that merges each speaker's consecutive lines into paragraphs, with the original timestamped transcript kept below it.
• Russian — Pick Russian in Settings → Transcribe and summaries, titles and chat come back in Russian instead of defaulting to English. Covered by auto-detect too.

Fixes

• macOS no longer shows an unexplained local-network prompt on first launch.
• FaceTime and Phone.app calls now trigger the meeting prompt like any other call.
• Interrupted Parakeet model downloads resume instead of starting over.
• An interrupted batch transcription keeps what it managed to transcribe instead of discarding the run.
• The same person is no longer split into two speakers when their voice bleeds across both channels.
• Quiet microphone speech is recovered in stereo recordings instead of being dropped.

Under the hood

• Diarization is roughly 6.65x faster, using Sortformer's offline chunk preset.
• Processing shows live per-stage progress while speakers are worked out, instead of one long opaque step.
• The Overview marks which notes still have their original audio, so you can see what can be re-transcribed.
• mlx is now pinned. It arrived as an unpinned transitive dependency, so a build could silently pick up a version that breaks Parakeet; the build now refuses to ship an unverified one.

Known limitations

• Speaker separation is macOS-only. Windows keeps the existing [You] / [Others] channel labels.
• Speaker accuracy depends on recording quality — overlapping speech and shared microphones are the hardest cases.
• Selecting Russian pins the language of the summary, title and chat. Transcription accuracy for Russian on the default Parakeet engine has not been measured; switch to Whisper in Settings → AI if the transcript itself is poor.

Upgrading

No action needed. Speaker separation is on automatically on macOS; speaker identification stays off until you enable it.

Thanks to our contributors

@Optic00, @valentinweyer, @ferebee, @catinapoke and @audreyt.

Downloads

• Apple Silicon (M1-M5): stenoAI-macos-arm64.dmg

The .zip and .yml assets below are used by the in-app auto-updater - no need to download them manually.

First Time Setup

1. Download the DMG
2. Install by dragging Steno to Applications
3. Launch app - setup wizard runs automatically
4. Grant microphone permissions when prompted
5. Start recording meetings!

Requirements

• Apple Silicon Mac (M1 or later) - Intel Macs are not supported (since v0.4.0)
• macOS 14.4 (Sonoma) or later
• Internet connection for initial setup
• Microphone access permissions

Intel Mac users

Steno is Apple Silicon only (since v0.4.0). Intel users should stay on v0.3.8, the last release supporting Intel Macs. Auto-update on existing Intel installs is gated and will not offer newer versions to those machines.