Speech engines & models
Choose between local Parakeet and Whisper recognition, and manage their model downloads.
On this page
Choose an engine#
Open Settings → Transcription → Engine. Leave it on Auto unless you want to compare engines or have a specific reason to force one.
| Setting | Recognition choice | When it is useful |
|---|---|---|
| Auto | Parakeet for a supported selected language; Whisper otherwise. | A practical default that follows your language selection. |
| Parakeet | Forces NVIDIA Parakeet TDT 0.6B v3. | CPU recognition for its 25 supported languages. |
| Whisper | Forces the selected Whisper model. | Broader language coverage or comparing recognition results. |
Engine Auto and language Detect automatically are separate settings. With Auto engine and Detect automatically language, TawkType selects Whisper.
Understand the resource cost#
| Parakeet TDT 0.6B v3 | Whisper LargeV3Turbo | |
|---|---|---|
| Processing | CPU only | Vulkan GPU when available, with a CPU fallback |
| Model storage | Around 0.7 GB download / roughly 640 MiB on disk | Around 1.6 GB / roughly 1.5 GiB on disk |
| Working memory | The project measured about 850 MB for the loaded app | The project measured about 1.5 GB VRAM with GPU use; CPU use needs more system RAM |
| Hardware preference | A modern multi-core CPU | A current graphics driver and roughly 2 GB free VRAM help |
Sizes are approximate; decimal GB and binary GiB displays differ. Measurements describe one development PC, not a guarantee for yours. Model loading, clip length, other programs and driver support all affect speed.
Parakeet chooses a bounded number of CPU threads based on your logical processor count. Whisper does not require the CUDA Toolkit: the application uses its Vulkan runtime when supported.
Download a model#
- Choose the engine and language in Settings → Transcription. If using Whisper, choose the Whisper model as well.
- Click Save, then reopen Transcription so its model status reflects your saved choices.
- In the download picker below Downloaded models, select the model that matches those choices.
- Click Download and wait for completion. Use Cancel if needed.
- Try a short dictation, or use guided setup to check that the model loads successfully.
Guided setup downloads and loads the selected model in one workflow. The normal dictation path does not silently start a large missing-model download. If it reports a missing model, return to setup or Transcription to request it.
Models live under %LOCALAPPDATA%\TawkType\models\. Changing the engine or language can require a different model; having one downloaded does not mean all combinations are ready. When you explicitly select Parakeet, the language picker is disabled because Parakeet detects its supported languages itself.
Whisper model choices#
LargeV3Turbo is the default Whisper selection. The app exposes the model names supported by its Whisper library, including smaller Tiny, Base and Small families and larger Medium/Large variants. English-only model names end in En; choose a multilingual model for other languages.
Smaller models use less storage and can run faster, but may recognize your speech less accurately. Compare them with the same microphone and a few representative passages. A model choice is a tradeoff, not an assurance of better results on every language.
Remove downloaded models#
The Downloaded models list shows files and storage usage. Select an entry and use Delete selected…, or use Delete all… to remove the listed model downloads. Read the confirmation before proceeding.
Removing the model needed by your selected configuration makes that configuration unavailable until you download it again. Model deletion is separate from deleting history or vocabulary. Avoid model maintenance in the middle of a dictation.
Compare without changing your words twice#
Use the same short passage with each engine and inspect the Heard text in History. This separates recognition quality from vocabulary and optional rewriting. Choose Literal or disable optional Claude rewriting for a clearer comparison, while remembering that local vocabulary rules still apply.
History’s Clean again action reruns text cleanup; it cannot replay a recording through a different speech model. To compare recognizers, make a new recording.