SnipVoice v1.2.0

SnipVoice v1.2.0 makes a recording's audio files explicit, builds the final audio about 10x faster, fixes the microphone boost, and lowers SnipVoice's memory and CPU use while idle.

What's changed

• Library > Files separates the final audio from the original (raw) tracks and shows the format and size of each.
• Download final audio saves a copy of the audio used for playback, in its original format. It no longer needs the original tracks.
• Download each original track (microphone or system audio) on its own, as MP3.
• Remove original tracks frees most of a recording's space: they are uncompressed, often around 30 MB per minute. It removes both tracks together and is only offered once a final audio file is saved. The confirmation warns that the recording can no longer be transcribed again. Afterwards, playback and downloads use only the final audio, and the transcript, summaries, chat, and citations stay. Automatic retention in Settings > Privacy follows the same rule.
• Original tracks are no longer pre-selected for removal, and removal never falls back to every track when none is chosen.
• Automatic clean-up: Settings > Privacy can remove original tracks after a number of days you choose (30 by default when turned on). SnipVoice applies it at startup and every six hours, only to recordings with a saved final audio and a finished transcript. The previous raw-audio policy was saved but never applied.
• Settings > Privacy in plain language: every option is a sentence with a short hint instead of internal codes. "Automatically move recordings to the trash after N days" now actually runs in the same pass, and moves recordings to the trash, where they can still be restored. Nothing is purged automatically.

Faster, better final audio

• About 10x faster: after a recording, the final audio is ready in about 20 seconds instead of 3-5 minutes for a 24-minute meeting. Adjust volume and audio downloads get the same speed-up.
• Mic boost that holds up: one desk knock or laugh no longer switches the boost off for the whole recording. Loud moments where you and the call overlap are limited instead of clipped.
• Balanced against the call: when the call is louder than you, your voice is raised to at least the call's speech level (up to 8x).
• Adjust volume rebuilds the MP3 in place instead of leaving "(2).mp3" copies.

Lighter in the background

• One model by default: recordings are transcribed with your dictation model unless you choose another in Settings > Recording ("Same as dictation").
• Memory: the dictation model (about 0.8 GB for the default) is released after 10 minutes without use. Quick follow-ups stay instant; the first dictation after a quiet spell waits about a second while it reloads.
• CPU: idle use while the window is closed dropped from about 1.3% to 0.3% of a core.

Downloads

• Windows x64: portable ZIP and per-user installer.
• macOS 14.4+ ARM64: ad-hoc-signed app bundle ZIP.
• FFmpeg source and build-information archives are included for the clean LGPL audio runtime.

See the full changelog for details.