Change the whisper model tier for accuracy vs speed

The Copy Assist panel uses a whisper.cpp model to transcribe received audio. Larger models produce more accurate transcriptions but use more VRAM/RAM and process audio more slowly. This page shows you how to switch between model tiers to balance accuracy against speed.

Before you start

• AetherSDR must be connected to a FLEX-8600 radio.
• Copy Assist must be visible (open it with View > Copy Assist or press Ctrl+Shift+T).

Steps

1. In the Copy Assist panel, click the Settings button (gear icon). This opens the Copy Assist settings dialog.
2. In the dialog, locate the Model tier combo box.
3. Click the combo box and select one of:
4. tiny — Fastest, lowest accuracy, least memory used.
5. base — Reasonable speed with moderate accuracy.
6. small — Good accuracy, noticeably slower.
7. medium — Highest accuracy, slowest performance, most VRAM/RAM required.
8. Close the settings dialog. The new model tier takes effect immediately after the underlying engine reloads.

What each control does

Control	Default	Valid range	Behavior
Model tier combo box	tiny	tiny / base / small / medium	Selects the whisper model size. Larger models improve accuracy but increase latency and memory usage.
Compute device combo box	GPU (CUDA/Metal)	GPU / CPU	Selects whether whisper runs on GPU (faster, needs VRAM) or CPU (slower, works everywhere).
Backlog indicator	0.0s	—	Seconds of received audio not yet transcribed. Colour escalates amber→red as backlog grows.

Choosing a compute device

By default, AetherSDR runs whisper on the GPU if one is available. You can force CPU or explicitly select a specific GPU in the Copy Assist settings dialog.

Steps

1. Open the Copy Assist settings dialog via the Settings button.
2. Locate the Compute device combo box.
3. Choose:
4. GPU (CUDA/Metal) — Uses a CUDA or Metal GPU. Faster, but requires sufficient VRAM.
5. CPU — Runs entirely on the CPU. Slower, but works on any system.

Note: On macOS, the first GPU query may take several seconds because the Metal shader library is compiled on demand. This happens in the background and does not block the Copy Assist panel.

Context-carry behavior

The Context-carry toggle in the Copy Assist panel continues the decoded prompt across the transcript. This is only available with the whisper backend; when using Sherpa or remote backends, the toggle is disabled.

Tips

• If you notice the Backlog indicator climbing (amber→red), switch to a smaller model tier or to CPU to help the engine catch up.
• The model files are downloaded automatically when you switch to a tier for the first time. Expect a brief delay and internet connection required for the initial download.
• If transcription is consistently inaccurate, try a larger model tier first before changing the compute device.
• When the Backlog indicator stays high on GPU, consider switching to CPU—some systems have GPU memory pressure even when the GPU is nominally available.

Related

• Copy Assist — Speech to Text overview
• Choose GPU or CPU for speech recognition
• Enable speech-to-text transcription on a slice