Changelog

Packaging and release-pipeline maintenance.

Meeting auto-capture (Windows, beta): a background helper detects an in-progress call — Teams, Zoom, Webex, Meet in a browser, plus WhatsApp/Telegram/Signal/Slack/Discord voice calls — records system audio and your microphone, and queues the result for transcription when the call ends. Detection is entirely local; no meeting APIs, no cloud.

Bundled with the Windows app: the native build runs the watcher in-process — no separate install, no scheduled task. A tray icon shows status (off / idle / recording) and can pause auto-capture.

Live recording indicator with elapsed timer and detected app, and the finished transcript opens automatically once it is ready.

Captures are written at 16 kHz mono, the rate Whisper transcribes at — a 2 h meeting is ~230 MB instead of ~700 MB.

Auto-captured meetings now use your actual settings. The watcher used to hard-code diarization and the small model, so every recorded meeting was diarized regardless of the Speakers toggle.

Long meeting uploads no longer time out at 60 s, and orphaned capture scratch files are cleaned up on watcher start.

Floating queue widget: active jobs surface in a bottom-right pill; click to expand a panel listing every non-terminal job with per-row cancel and click-to-attach.

Enqueue while running: the drop zone and URL field stay usable during transcription, so more files can be queued without waiting.

Diarization progress is real: pyannote's step events (segmentation → embeddings → clustering) now drive the bar instead of it freezing at 82%.

Cancel is now real for queued and running jobs, including during the URL download phase and around the diarization pipeline.

Diarization compatibility with pyannote.audio 3.4+ (#24) and with huggingface_hub 1.0+ / torch 2.6+. Thanks to @Tiritibambix.

Export filenames with non-latin-1 characters (#25) no longer fail — exports emit RFC 5987 headers.

New GET /api/jobs lists non-terminal jobs with status, progress, queue position and creation time.

Benchmark elapsed time: each model result includes load + inference combined, with total wall-clock time for the run.

Benchmark tool: runs tiny/small/medium against a standard 11 s clip and reports load time, inference time and real-time factor for each. Results can be shared to the community with a pre-filled GitHub issue.

Recording alias: Rename any recording with a display name independent of the source filename. Used as title in Markdown exports.

Transcript tab decluttered: Collapsed inner tab bar, search bar, and segment selection bar into a single compact toolbar.

AI Analysis slide-over: Moved from a sub-tab to a slide-over panel (lightbulb icon), keeping the transcript always visible.

Bulk speaker assignment: Select multiple segments and assign a speaker name in one action via the contextual toolbar.

AI result markdown rendering: AI Analysis output now renders as formatted markdown instead of plain text.

Enhanced Markdown export: Speaker runs merged into paragraphs, timestamp only at start of each speaker turn, metadata header with duration, language, speaker list, and date.

Bulk Markdown export: Multiple recordings export as a single combined file with a table of contents and separators.

Fix torch/torchaudio dependency caps for Python 3.13 and lightning compatibility.

Error in transcription.py #22 (thanks @nyfon).

Fix GPU release build.

Create distributable with GPU enabled.

CORS restricted to localhost: Prevent cross-origin requests from arbitrary websites.

Exit endpoint CSRF token: /api/exit now requires a per-session token, blocking DNS-rebinding attacks.

Audio path bounds check: Path traversal prevention in /api/audio/{job_id}.

Frontend XSS fix: Folder/tag menus rebuilt with DOM API, eliminating injection via names containing ' or </script>.

Chunked file upload: 1 MB streaming chunks prevent OOM on large audio files.

Atomic settings write: Writes to .tmp then renames atomically, preventing corrupt settings on crash.

9 new test files added covering exports, settings, search escaping, session lifecycle, and more.

Microphone recording: Record directly from your microphone with pause/resume support and a live timer. Auto-queued into the batch transcription flow.

README: badges, competitor comparison table, Telegram community link, roadmap section.

CONTRIBUTING.md and GitHub issue templates added.

Feedback link in sidebar footer.

URL import from YouTube, TikTok, Instagram, Facebook, X, Vimeo, and Twitch via yt-dlp.

Automatic platform tagging on URL imports (e.g., youtube, tiktok).

Backend API modularized into dedicated router modules under backend/api/routes/.

Centralized status strings in backend/core/messages.py.

Unit tests for diarization, audio normalization, Whisper model cache, CUDA/VAD error classifiers.

Transcription pipeline split into focused modules under backend/core/.

Worker replaced with asyncio background task using asyncio.Queue.

New TranscriptionConfig model with configurable compute_type, device, vad_filter, word_timestamps, and more.

DB indexes added for frequently queried fields.

Update check: Frontend displays a banner when a newer GitHub Release is available.

Google Colab integration: Offload transcription to a free Colab GPU while keeping files local.

Bulk actions: Select multiple recordings and apply bulk move, tag, or delete operations.

Batch directory import: Drop a folder of audio files for batch transcription.

Mobile UI: Off-canvas sidebar overlay, always-visible segment action buttons on touch devices.

Docker: Compose split into base, dev override, and production overlay. Traefik HTTPS support.

AI Analysis Engine: Per-recording LLM-powered analyses — summaries, action items, translation, custom prompts — with streaming results.

LLM Settings: Configure base URL, model name, and API key from the UI. List and pull models.

Drawer sidebar, inner tab panels, client-side action logs, and Help modal.

Waveform player with interactive seeking and segment highlighting.

Collapsible console log panel with timestamps.

Multiple file upload; video file support via ffmpeg audio extraction.

macOS release added.

Global search with live filtering.

Server-side Hugging Face token persistence.

Automated platform-specific ffmpeg download on first run.

Fixed 10-color palette for tags and folders with server-side validation.

Per-folder and per-tag counts in sidebar.

Tag-click filtering scoped to selected folder.

Initial stable release.

PyInstaller packaging fixed for speaker diarization.

Windowed runtime crash fixed during diarization.