Transcription & AI

Autorec transcribes recordings locally using whisper.cpp and optionally generates AI summaries via any OpenAI-compatible API.

How Transcription Works

1. After a recording finishes, autorec extracts the audio track. This happens on its own when Auto-transcribe recordings is on (it is off by default); otherwise start it from the prompt after the recording or from the library.
2. The audio is processed by the selected Whisper model entirely on your machine
3. Two output files are created alongside the video:
   • .txt — plain text transcript
   • .srt — subtitle file with timestamps (can be turned off in Settings)

No audio or video data leaves your computer during transcription. It runs on the GPU when one is available (Vulkan on Linux and Windows, Metal on macOS) and on the CPU otherwise; Force CPU in Settings skips the GPU.

Languages

Autorec transcribes 17 languages: English, German, Spanish, French, Italian, Portuguese, Dutch, Polish, Swedish, Turkish, Russian, Ukrainian, Arabic, Hindi, Chinese, Japanese, and Korean. Autorec detects the spoken language from the first seconds of audio, so most calls need no setup. If your calls are always in one language, set Spoken language under Settings > Transcription to it. A pinned language avoids a wrong guess on short or noisy audio, or on a call that opens with a few words in another language.

All five models understand every one of these languages. For anything other than English, medium or large is noticeably more accurate than the smaller sizes.

Whisper Models

Models are not downloaded automatically; you pick one and download it in Settings. They are stored in ~/.local/share/autorec/models/ (Linux and macOS) or %LOCALAPPDATA%\autorec\models\ (Windows).

Model	Size	Speed	Accuracy	Best For
tiny	~75 MB	Fastest	Basic	Quick notes, low-power machines
base	~142 MB	Fast	Good	Default — recommended for most users
small	~466 MB	Moderate	Better	When accuracy matters more than speed
medium	~1.5 GB	Slow	High	Non-English languages, difficult audio
large	~3 GB	Slowest	Best	Maximum accuracy, powerful hardware

Downloading Models

1. Open Settings, the gear icon in the autorec window
2. Go to the Transcription section
3. Select a model size
4. Click Download — the model downloads once and is reused for all future transcriptions

AI Summaries

AI summaries use a cloud API to generate a title and summary from the transcript text. Only the text is sent — no audio or video.

Setup

1. Open Settings > AI Summary
2. Pick a Provider, which fills in the base URL and suggests models, or choose Custom and enter the API Base URL (e.g., https://api.openai.com/v1)
3. Enter your API Key
4. Choose a Model (the default is gpt-4o-mini)
5. Turn on Auto-summarize after transcription

The Summary prompt is editable, so you can ask for decisions, action items or deadlines instead of a general summary.

Compatible Services

Any service with an OpenAI-compatible chat completions endpoint works:

• OpenAI — https://api.openai.com/v1
• OpenRouter — https://openrouter.ai/api/v1
• Local models (Ollama, LM Studio, etc.) — use your local endpoint

What Gets Generated

For each transcribed recording, autorec generates:

• Title — a short, descriptive title for the meeting
• Summary — a concise summary of key points discussed

Both appear in the video library and the video detail view, making it easy to find the meeting you need without rewatching.