v4.0.0 - 2-Pass Autofix, Speaker Diarization & OpenRouter Gateways

Highlights in v4.0.0

• 2-Pass Hybrid Autofix (--autofix): Combines Whisper-1's frame-accurate timecodes with intelligent LLM context cleanup to correct technical jargon, domain terms, brands, and acronyms in place.
• Automatic Speaker Diarization: Detects and labels speaker turns ([Speaker 1]: ..., [Speaker 2]: ...) during autofix (disable with --no-diarize).
• OpenAI-Compatible Gateways & Models (Issue #1): Full support for OpenRouter, Groq, LiteLLM, vLLM, Deepgram via --model and --base-url.
• Single-Token Auto-Detection: Works seamlessly with just OPENAI_API_KEY (defaults to gpt-5.6-luna) or OPENROUTER_API_KEY (defaults to google/gemini-3.7-flash). Auto-mixes if both keys are set!
• Zero Audio Speedup Distortion: Removed 1.2x speedup filter across all files to preserve 20–50ms acoustic transients and eliminate phonetic blur on foreign terms.

v3.8.0 - macOS Quick Actions & Productivity Integrations

Added

• macOS Finder Quick Action: Run npx @illyism/transcribe --install-mac-action to add a native right-click "Transcribe Subtitles" menu option in macOS Finder with system notifications.
• macOS Productivity Integrations: Added setup guides in README for Apple Shortcuts, Raycast/Alfred, and Automator Drop Zone folders.
• Unit Tests: Added a bun test suite (103 tests) covering URL vs path detection, input routing, media discovery, time parsing, and SRT timeline math.

Fixed

• URL vs Path Disambiguation: Improved URL parsing in remote providers to prevent local filesystem paths from being misidentified as URLs.
• SRT Millisecond Rounding: Fixed floating point truncation that could shift a cue by 1ms.
• Negative Timestamps: Clamped negative --offset timestamps to 00:00:00,000.

npm

v3.7.0 - Instagram Reels, Browser Cookies & Agent Skills

Added

• Instagram Reels Support: Transcribe Instagram Reels (/reel/, /reels/, /p/) directly using automatic browser cookie login resolution.
• --cookies-from-browser Flag: Pass --cookies-from-browser <browser> (e.g. chrome, safari) to override browser cookie detection for authenticated sites.
• Agent Skills: Published user-focused AI agent skills under skills/ (compatible with skills.sh) for transcribe, summarize-transcript, video-to-social, youtube-chapters, and format-converter.

Refactored

• Modular Remote Architecture: Split monolithic src/youtube.ts into src/remote/ module (providers.ts, urls.ts, ytdlp.ts, cookies.ts, progress.ts).

npm

v3.6.0

Changed

• Short files stay raw: Speed optimization (1.2x) only runs for media 5 minutes or longer; shorter files use original audio automatically.
• Quieter CLI output: Less noise for single-chunk jobs, silent temp cleanup, and a single saved-line with language/duration in folder mode.

npm

v3.5.0

Added

• Bulk Folder Transcription: Pass a folder path to interactively select all or choose individual media files, then transcribe them in bulk. Each .srt is written next to its source file (or into --output).

v3.4.0

Added

• Screen Studio Support: Transcribe .screenstudio recordings directly — extracts microphone audio from HLS segments and supports both bundle directories and zip archives.

Changed

• Always-on Chunking: All transcriptions are now split into ~20 minute chunks by default for reliability (override with --chunk-minutes).
• Screen Studio Uses Raw Audio: Screen Studio inputs skip speed optimization automatically to preserve original timing.
• Build Fix: Mark openai as an external dependency during bundling to fix Whisper multipart upload errors in the published CLI.

v3.3.0

Added

• Universal URL Support: Support for transcribing any media URLs supported by yt-dlp (including X/Twitter, and more), expanded from just YouTube.
• Honest Progress Bars & Live Stats: Real-time progress bars or live activity lines with byte-counts and speed, adapting automatically for HLS and external FFmpeg downloads.

v3.2.0

Summary

• Transcribe long media chunks with up to 8 concurrent Whisper API requests.
• Precompute chunk durations and offsets before transcription so merged SRT timestamps stay deterministic.
• Normalize npm package metadata for publishing.

Verification

• Published @illyism/transcribe@3.2.0 to npm.
• Verified build and npm publish dry-run before release.

🚀 v3.1.0 - Long Movies & Workflow Improvements

This release focuses on making transcription more reliable for long-form content and providing better controls for video editing workflows.

✨ Highlights

• 🧩 Automatic Chunking: Media files over 45 minutes are now automatically split into 20-minute chunks. This significantly improves reliability with the OpenAI Whisper API and ensures stable processing for 2+ hour movies.
• 🕒 Timecode Offsets: Use the new --offset flag to shift timestamps (e.g., --offset 01:00:00.000 or --offset 3600). Perfect for syncing with existing editorial timelines.
• 📂 Custom Output: Specify exactly where you want your SRT files with the --output (-o) flag. Supports both specific file paths and directories.
• 🎙️ Optimized Extraction: Audio extraction now uses mono and 16kHz settings by default, reducing file size by up to 99% while maintaining 100% dialogue clarity.

🛠️ New CLI Flags

• --output / -o: Custom SRT path or directory.
• --offset: Shift subtitle timestamps (supports seconds or HH:MM:SS.mmm).
• --chunk-minutes: Force custom chunk duration for long inputs.

🐛 Improvements

• Enhanced FFmpeg error reporting with actionable guidance.
• Improved cleanup logic for temporary chunk files.
• Updated README with detailed usage for long-form content.

v3.0.3 - Critical Timestamp Fix

🐛 Critical Bug Fix

Fixed

• Critical: Fixed timestamp adjustment direction when using speed optimization
  • Timestamps were being divided by speed factor instead of multiplied
  • This caused SRT files to be 17% shorter than original audio (e.g., 40min audio had timestamps ending at 28min)
  • Now correctly multiplies timestamps by speed factor to match original audio duration
  • Example: 40:17 audio → sped up to 33:34 → Whisper timestamps correctly adjusted back to 40:17

📦 Installation

npm install -g @illyism/transcribe@3.0.3
# or use directly
npx @illyism/transcribe@3.0.3 video.mp4

Links

• 📦 NPM Package
• 📚 Full Changelog