Whisper Models Compared: Audio Note Model Guide
Which Whisper model to use in Audio Note: Tiny to Large-v3 Turbo, English-only, Chinese community, and quantized variants, plans, and testing your own audio.
Short answer
- Not sure: start with Small.
- Want better accuracy on a decent computer: Large-v3 Turbo.
- Older computer or just a quick look: Tiny or Base.
- English only: compare the English-only variant with the multilingual one.
- Mostly Chinese: try the Large-v3 Chinese model, or non-Whisper models such as SenseVoice and Qwen3-ASR.
Official models
File transcription offers several Whisper sizes. Sizes come from the app's model configuration:
| Model | File size | Good for |
|---|---|---|
| Tiny | 80MB | Quick previews, older computers, checking that audio transcribes |
| Base | 150MB | Short audio, quick notes |
| Small | 500MB | Starting point for meetings, classes, interviews |
| Medium | 1.5GB | Accuracy matters, harder audio |
| Large-v2 | 2.9GB | Consistency with older results |
| Large-v3 | 3.1GB | Quality first, material you keep |
| Large-v3 Turbo | 1.6GB | Near large-model results, faster |
Tiny, Base, Small, and Medium each also have an English-only variant of the same size.
File size is not runtime memory or VRAM. Speed also depends on your computer, acceleration, threads, audio length, and quality.
Community models
Pro also includes community models:
| Model | File size | Notes |
|---|---|---|
| Large-v3 Chinese | 3.1GB | Chinese community fine-tune, best Chinese accuracy |
| Large-v3 Turbo Q5 | 547MB | Quantized Large-v3 Turbo, much smaller |
| Distil Small / Medium / Large-v2 English | 336MB to 1.52GB | Distilled English models, faster |
| Distil Large-v3 | 1.52GB | Distilled Large-v3 (downloads distil-large-v3.5, English only) |
Quantized (Q5) models store weights at lower precision: smaller files and memory use, possibly slightly lower accuracy. Compare on your own audio.
Sources: Large-v3 Chinese comes from BELLE-2/Belle-whisper-large-v3-zh; Large-v3 Turbo Q5 from the whisper.cpp model repo; the Distil models from distil-whisper.
Which plan includes what
| Plan | Whisper models | GPU acceleration |
|---|---|---|
| Free | Tiny, Base, Small (with English variants) | No |
| Standard | All 11 official models | Yes |
| Pro / Lifetime | Official plus all community models | Yes |
See the pricing page and in-app messages.
How to use each size
Tiny for quick checks: not accurate enough for final copy, but light enough to confirm the audio and language.
Base for light work: fine for clean, short audio; move up to Small or Medium for noise, accents, or jargon.
Small is the default for most people: meetings, classes, interviews. The trade-off is missed detail on hard audio.
Medium for accuracy: noisy recordings and multi-speaker conversations, at the cost of speed and resources.
Large-v3 for quality: material you transcribe once and keep, such as interview archives or research.
Large-v3 Turbo is often the better large model: about half the size of Large-v3 and faster.
Let the app recommend for your computer
The home screen has a Performance Benchmark. It runs a roughly 35-second English sample through Whisper Base and a lightweight realtime model, records speed, error rate, and CPU and GPU load, and combines that with your hardware (chip, GPU, VRAM, memory):
- Acceleration: Metal on Apple Silicon, CUDA on NVIDIA, Vulkan on discrete AMD GPUs; CPU on Intel Macs, Windows machines with only integrated graphics, and virtual machines. If the GPU turns out slow in the test, it falls back to CPU.
- Whisper model for files: it picks from Tiny, Base, Medium, and Large-v3 only. Medium or Large-v3 is recommended only when both tests look strong on a GPU (more VRAM, memory, or a newer chip leans toward Large-v3). Most machines get Base; machines that struggle get Tiny.
- Whisper or a lightweight model: if the lightweight model is clearly faster on your machine with a similar error rate, it becomes the file recommendation. For live transcription it works the other way: Whisper is recommended only when it is fast, accurate, and leaves CPU headroom; otherwise a lightweight realtime model.
Click Apply Recommended Settings to use the result. Model lists then mark the recommended model and flag Whisper models heavier than your machine handles well. If the recommended model is not in your plan, pick the closest one that is. Treat it as a starting point and test with your own audio.
Hardware
Apple Silicon Mac (M1 and later). Start with Small or Large-v3 Turbo. If you process long recordings, test one sample locally before committing to Medium or Large-v3.
Windows with an NVIDIA GPU. Enable GPU acceleration on a paid plan and choose CUDA or Vulkan. CUDA depends on driver and card; see NVIDIA GPU and CUDA compatibility.
No discrete GPU. Prefer Tiny, Base, or Small. Larger models run but may be slower than you want. On Standard or above, SenseVoice or Moonshine (both non-Whisper) are fast on CPU.
More in the GPU transcription docs.
Benchmarks
We have already run the Whisper sizes against the other local models on one clip set; the full data is in Local Speech Recognition Models Tested. Your own recordings are still the most reliable comparison.
Compare on your own audio
Do not apply a WER number from the internet to your situation. Accents, noise, microphones, bitrate, and jargon all matter.
- Pick a 2 to 3 minute recording typical of your work.
- Transcribe it with Small, Large-v3 Turbo, and Large-v3.
- Count errors in the first minute the same way each time.
- Time each run and divide by the audio length to get your real-time factor.
- Pick the one whose errors and wait you can live with.
Multilingual vs English-only
If the audio is not purely English, use a multilingual model. For English only, compare the English-only and multilingual variants.
For Chinese, also try the non-Whisper models: SenseVoice is fast and good on Chinese (Standard and up); Qwen3-ASR is more accurate but slower (Pro). See the models overview.
FAQ
Where are models stored? In the app's model folder; manage and delete them in settings.
Do old transcripts change when I switch models? No. Finished transcripts stay as they are.
Why does a bigger model sometimes add text? Occasionally large models "fill in" text during silence or noise. Try VAD settings or compare another size. See the per-engine parameter docs.
