Settings → Speech Engines
Speech Engines is where you choose and install what turns speech into subtitles. You can have several and pick one for each transcription.
The three engines
| Whisper | Exact timing to the word and the widest choice: 99 languages and models from Tiny to Large. It can repeat or invent text on music and silence, which the clean-up options work against. Runs on the CPU, or the graphics card with a CUDA, Vulkan or Metal build (chosen on Dependencies). |
|---|---|
| Qwen3-ASR | Strong on Chinese and its dialects, accented English and about 30 languages, which it detects itself. About 0.9 GB, CPU only, Windows only. Each subtitle is timed by the stretch of speech it came from, not word by word. Uses a vocabulary hint. |
| Parakeet | Fast, accurate English and 24 other European languages, detected automatically, with punctuation and capitals built in. About 0.5 GB, CPU only, Windows only. No Turkish, Arabic or Asian languages, no vocabulary hint, and each subtitle is timed by its stretch of speech. |
For subtitles timed exactly to the word, or any language, use Whisper. For Chinese, dialects and accents, try Qwen3-ASR. For fast, accurate English or European languages, try Parakeet. Qwen3-ASR and Parakeet need about 1.2 GB of memory per running file.
Installing and choosing an engine
Each engine has a card with a status — Ready, Needs a model (Whisper without a downloaded model), Not installed or Not available on this system — and buttons: Install, Use by default, Manage (for Whisper, which sends you to Dependencies to pick a build), Update and Remove. An engine with a newer pinned version shows Update available. The model list on Transcribe also offers an engine you have not installed, with its size; choosing it installs it and selects it when done.
Whisper models
| Tiny | 75 MB — Fastest |
|---|---|
| Base | 142 MB — Fast |
| Small | 466 MB — Medium |
| Medium | 1.4 GB — Slow |
| Large (V3) | 2.9 GB — Slowest, best quality |
| Silero VAD | 2.5 MB — powers the Voice Activity Detection option in General settings; not a transcription model |
Larger models are more accurate but slower. Each row shows download progress, then Installed or Not installed, with Download, Cancel, Retry or Remove. Whisper itself must be installed first, on Dependencies. If you copied model files into the models folder yourself (ggml-*.bin), click Refresh model list to detect them; the folder path is shown at the bottom of the screen.