Blab 0.7.11

What's Changed

• docs: rewrite the README for everyone by @jurecerkez-code in #45
• Transcripts say what was said: per-stretch language, honest loop warning, silence detector context by @jurecerkez-code in #46
• chore: release 0.7.11 by @jurecerkez-code in #47

Full Changelog: v0.7.10...v0.7.11

Blab 0.7.10

Same app, cleaner code. Nothing you can see or hear changed: this release is identical to 0.7.9 in everything the app does.

The repository got its cleanup: no dashes left in comments or docs, one redundant check removed, stale branches and release leftovers gone.

Windows, Mac and Linux installers, all three models inside, as before. Free, offline, no account.

Blab 0.7.9

What's Changed

• Quiet mics and dense speech no longer lose their words by @jurecerkez-code in #41

Full Changelog: v0.7.8...v0.7.9

Blab 0.7.8

What's Changed

• Copy all gives pure text by @jurecerkez-code in #40

Full Changelog: v0.7.7...v0.7.8

Blab 0.7.7

Full Changelog: v0.7.6...v0.7.7

Blab 0.7.5

Fixed

Re-transcription was completely broken in 0.7.3 and 0.7.4: every chunk starts with timestamp tokens, the filter removed them all, and the decoder crashed on the empty input (“token_ids must be a non-empty array of integers”). Timestamps are now filtered by token id range, an all-timestamp run degrades safely to the end token, and the broken regex no longer eats words containing digits. Three regression tests cover it.

The empty-array crash, the broken regex from 0.7.3, and the memory fix from 0.7.4 are all in this release.

Blab 0.7.4

Fixed

A failed transcription now drops the engine worker, so the next attempt starts with a clean memory heap. This fixes the "failed to allocate a buffer" error that could appear when re-transcribing a long recording with the Best model.

Blab 0.7.3

Blab 0.7.3

Fixes one genuinely ugly thing: the timestamp junk.

• No more <|1.32|> tokens in transcripts. Whisper emits those tokens inline, and the engine did not mark them as special, so they leaked into the in-progress transcript and the live captions, and even made the incremental decoder repeat words. They are now filtered at the token level before decoding, with a defensive strip over the final text and segments.
• Clean transcript view. The Transcript block shows only the words. Times still live on the highlights and your notes, and every line is still clickable to play from that moment.
• Everything else from 0.7.2 unchanged: all three models inside, out of the box, offline forever.

53 tests green.

Blab 0.7.2

Blab 0.7.2

Ready to use out of the box.

All three models inside the installer. Fast, Balanced and Best ship with Blab itself. Install once, and you are done: no accounts, no downloads, no configuration, works offline forever.

The right model for your machine, picked automatically. First launch chooses Best on Apple Silicon, Balanced on a modern laptop (8+ cores, 8+ GB RAM), Fast on older machines. The picker in the app is only there if you want to override it.

Everything from 0.7.0 and 0.7.1: silence-aware transcription (on-device voice activity detection), live captions, meeting capture (mic + computer audio), file import, SRT/VTT export, re-transcribe anytime, crash-safe transcripts, and diagnostics that actually hear your microphone.

Verified: 53 automated tests, end to end check green on Windows, all three model files validated by onnxruntime before packaging. Windows (this file); macOS universal DMG and Linux AppImage come from the same source. Expect one warning screen on install, it is normal for a free unsigned app.

Blab 0.7.1

Blab 0.7.1

Small, measured update on top of 0.7.0.

What's new

• Transcribe anytime. Every recording can be transcribed or redone from its detail view, with whichever model the picker currently has. Record with the fast model, upgrade to the balanced one later. Same file, no re-recording.
• A hard ceiling on runaway transcripts. Each 30 second pass is bounded to 224 tokens (whisper.cpp's proven --max-len limit), so a stuck Whisper can no longer produce a wall of repetition.
• Measured before choosing. Beam search and WebGPU were benchmarked and rejected for this app. The bundled transcription engine has no beam search, and WebGPU would require fp32 weights that triple the download. VAD and the model picker remain the accuracy levers.

Verified: 52/52 tests, app:check green, base model about 3.4x realtime on a laptop CPU. Windows (this file), macOS universal DMG and Linux AppImage come from the same source.