Model Selection Guide
StarWhisper uses OpenAI's Whisper AI models for transcription. Each model offers a different balance of speed, accuracy, and resource usage.
Default to Small on Auto
For typical dictation (short clips, conversational speech, English and most Latin-script languages), the Small model is the right default. Bigger is NOT always better: Medium and Large hallucinate more on short clips and can invent words when there is silence or background noise. Use Medium or Large only in the specific scenarios listed below.
Model Comparison
| Model | Size | Speed | Accuracy | RAM | Best For |
|---|---|---|---|---|---|
| Tiny | 75 MB | ~6x real-time | 85% | ~200 MB | Quick notes, low-power devices |
| Base | 142 MB | ~4x real-time | 90% | ~400 MB | General dictation |
| Small Recommended | 466 MB | ~2x real-time | 94% | ~600 MB | Daily use, professional work |
| Medium PRO | 1.5 GB | ~1.5x real-time | 96% | ~1 GB | Long-form continuous audio, non-Latin scripts (CJK, Arabic, Cyrillic) |
| Large PRO | 2.9 GB | ~1x real-time | 98% | ~1.5 GB | Batch file transcription only, not recommended for live dictation |
How to Choose
Start here: leave the app on Auto. Auto picks Small for typical dictation, which is the right call for the vast majority of users. The notes below are for the cases where you want to override the default.
Choose Small if (recommended default):
- You are doing typical dictation (emails, notes, chat messages, coding, short clips)
- You speak English, Spanish, French, German, Italian, Portuguese, or other Latin-script languages
- You want the best practical balance of accuracy and speed for live use
Choose Tiny or Base if:
- You are on a low-power laptop, tablet, or older PC with no GPU
- You want the fastest possible response and can accept slightly more errors
- Storage space is very limited (Tiny is 75 MB)
Choose Medium (Pro) ONLY if:
- You transcribe long-form continuous audio (meetings, lectures, podcasts) where the model has minutes of context to work with
- You dictate in a non-Latin script language (Chinese, Japanese, Korean, Arabic, Russian, Thai, Hindi, etc.) where Medium handles complex character sets better
- Note: on short dictation clips, Medium is more likely to hallucinate than Small. Test on your own audio before switching.
Choose Large (Pro) ONLY if:
- You are doing batch file transcription of recorded audio (not live dictation)
- You have time for slower processing and a GPU with enough VRAM
- Note: Large is NOT recommended for live dictation. It is slower, hallucinates more on short context, and rarely beats Small in real-world voice typing.
Model Switching
To change your model:
- Right-click the recording circle
- Select Settings
- Go to Transcription tab
- Select your model from the dropdown
- Click Download if needed
- Restart StarWhisper if prompted
Download Required
When switching to a new model, StarWhisper will download it (75 MB to 2.9 GB). This happens once per model. After download, no internet is required for local transcription.
GPU Acceleration Impact
With GPU acceleration enabled, all models transcribe significantly faster:
| Model | CPU Speed | GPU Speed | Speedup |
|---|---|---|---|
| Tiny | ~6x real-time | ~30x real-time | 5x faster |
| Small | ~2x real-time | ~20x real-time | 10x faster |
| Large | ~1x real-time | ~10x real-time | 10x faster |
Speed vs Accuracy Trade-off
The following table shows real-world performance for a 1-minute recording:
| Model | CPU Time | GPU Time | Words/min |
|---|---|---|---|
| Tiny | 10 seconds | 2 seconds | ~150 words |
| Small | 30 seconds | 3 seconds | ~140 words |
| Large | 60 seconds | 6 seconds | ~135 words |
Memory Usage
Consider your available RAM when choosing a model:
| Your RAM | Recommended Models |
|---|---|
| 4 GB | Tiny, Base |
| 8 GB | Tiny, Base, Small |
| 16 GB+ | All models including Medium and Large |
Recommendation Summary
Daily dictation (default)
Small on Auto. Best practical accuracy for short clips, emails, notes, chat, and coding. This is the right answer for almost everyone.
Low-power device
Base or Tiny. Choose when GPU is unavailable and CPU is constrained. Tiny stays usable even on old laptops.
Long-form or non-Latin scripts
Medium (Pro). Use for meetings, lectures, podcasts, or dictation in Chinese, Japanese, Korean, Arabic, Russian, etc. Not ideal for short live dictation.
Batch file transcription
Large (Pro). For processing recorded audio files where speed does not matter. Not recommended for live voice typing.
Why bigger is not always better
It is intuitive to assume Large = most accurate. In practice, for short dictation clips, Small frequently beats Medium and Large because the larger models try harder to "complete" the sentence, which makes them invent words on silence or background noise. Whisper itself documents this hallucination tendency. The safest bet for live typing is Small. Step up only when you have a clear reason listed above.