Speed and accuracy are two sides of the same dial. Weesper Neon Flow runs entirely on your computer, so the result depends on three settings you control:
- Profile — the size of the recognition model (Minimal → Professional)
- Quality level — the preset slider in Settings → Recognition Quality
- GPU acceleration — the toggle in Settings → Recognition Quality → GPU acceleration
This guide gives you a step-by-step tuning routine for both problems: when transcription feels too slow, and when accuracy is not good enough.
If transcription feels slow
Try these steps in order. Stop as soon as the speed is acceptable.
-
Switch to a smaller profile. This is the biggest factor by far. Move from Extended, Expert, or Professional down to Standard, Everyday, Compact, or even Minimal. Open Settings → Recognition Quality and pick a lighter profile.
-
Turn off GPU acceleration on older or integrated GPUs. Open Settings → Recognition Quality → GPU acceleration and disable it. On older laptops or machines with an integrated graphics chip, the CPU path is often faster than the GPU path for small models. The app’s own description says it: “Turn off if local processing feels slower or unstable.”
-
Lower the Quality level by one step.
-
Counter-intuitive tip — once you are on a smaller profile, try raising Quality back up. A smaller profile with a higher Quality level is often both faster and more accurate than a larger profile with a lower Quality level. The profile size is the dominant factor; once it is right for your hardware, you have headroom to push Quality up.
-
On modern machines, keep GPU acceleration on. Apple Silicon Macs (M1 and later) and Windows machines with a recent discrete GPU (NVIDIA, AMD) get a clear speed boost from GPU acceleration. Weesper uses Metal on macOS and Vulkan on Windows.
-
Close heavy applications during dictation, especially video calls, virtual machines, and large IDEs.
If transcription quality is poor
-
Choose the spoken language explicitly. Open the app, go to Speech Task (the language and mode page), and pick the language you will speak instead of leaving Auto-detect. Auto-detect works, but the app itself notes: “Transcription works best when the spoken language is set correctly.”
-
Raise the Quality level. In Settings → Recognition Quality, move the preset one step up.
-
Move up one profile size. For example, Standard → Extended, or Extended → Expert. Larger profiles are slower but noticeably more accurate. See Choosing the Right AI Model for the full comparison.
-
For domain vocabulary, use the Dictionary to add words, names, and acronyms that are misrecognized, and the Output Style for tone and structure. See Improving Transcription Accuracy for the complete list of accuracy tips.
Test one change at a time
The easiest way to find your sweet spot is to dictate the same 10-second sentence every time, change one setting, then redictate. Otherwise it is hard to tell which knob actually moved the needle.
A short test sentence covering numbers, punctuation, and a few technical words is ideal — it stresses both speed and accuracy in a realistic way.
Related
- Choosing the Right AI Model — Full comparison of profiles
- Improving Transcription Accuracy — Beyond settings: speaking style, microphone, dictionary
- Windows, ARM, GPU, and NPU performance — Windows-specific notes (Snapdragon, NPU)