Windows, ARM, GPU, and NPU performance

Why speed varies

Weesper Neon Flow processes your voice locally on your computer. This is excellent for privacy, but speed depends directly on your machine.

The main factors are:

• the selected recognition profile;
• the quality level;
• processor performance;
• available memory;
• the load from other applications;
• available hardware acceleration depending on your version.

A larger profile can be more accurate, but it requires more time and memory.

Fast mode is not real-time text

A profile called “Instant”, “Fast”, or similar means Weesper Neon Flow prioritizes speed. It does not mean the text appears while you speak.

The workflow stays the same:

1. you hold the shortcut;
2. you speak;
3. you release it;
4. Weesper transcribes;
5. the text is inserted.

Recommended settings if it is too slow

Try these steps in order:

1. Switch to a lighter profile.
2. Reduce the quality level.
3. Test with a 5 to 10 second dictation.
4. Close heavy applications.
5. Restart Weesper Neon Flow.
6. Check that your computer is not in battery saver mode.

If the lighter profile is much faster, your previous profile was probably too heavy for your setup.

Windows ARM and Snapdragon machines

Some recent Windows machines use ARM processors, for example Snapdragon X Elite.

In the current version, Weesper Neon Flow may not use the Qualcomm Hexagon NPU. It is therefore possible for Task Manager to show 0% NPU usage, even during transcription. This is not necessarily a bug: it simply means NPU acceleration is not used by this version.

GPU and hardware acceleration

Weesper Neon Flow uses Metal on macOS and Vulkan on Windows to accelerate transcription on the GPU. The toggle is in Settings → Recognition Quality → GPU acceleration and is enabled by default.

• On modern machines with a discrete GPU (NVIDIA, AMD), keep it on.
• On older laptops or machines with an integrated GPU, try turning it off — the CPU path is often faster for small models. The app’s own description confirms it: “Turn off if local processing feels slower or unstable.”

To compare fairly, dictate the same sentence, with the same profile and the same quality level, with and without acceleration.

For the full speed-vs-quality tuning routine, see Adjusting Speed and Transcription Quality.

When GPU Acceleration Silently Stays Off

On some machines — particularly laptops with an integrated GPU (for example Intel Iris Xe) — Weesper Neon Flow may keep transcription on the CPU path even though the GPU acceleration toggle is on and Vulkan is available. This is a deliberate compatibility policy for certain integrated GPUs, not a bug or a sign that acceleration is broken on your machine in general.

If you suspect this is happening on your computer:

• Toggle GPU acceleration off and on again, and compare a short, identical dictation both times.
• If timing is the same either way, your current app version is likely keeping this specific GPU on the CPU path by policy.
• Try a lighter recognition profile (“Fast”/“Instant”) for a same-sentence comparison — this has a bigger effect than the GPU toggle on integrated graphics.

If you would like this reported for your specific hardware, contact support with your exact GPU model and app version.

When to contact support

Contact us if a short dictation remains very slow with a light profile.

Include:

• your Windows version;
• your processor;
• your amount of RAM;
• the selected recognition profile;
• the quality level;
• the approximate duration of a 10 second dictation.