Copy Assist — Speech to Text¶

Copy Assist provides real-time speech-to-text transcription for the active slice, powered by whisper.cpp. It decodes received voice audio into a scrolling, color-coded transcript, helping you capture and review QSO content without relying on notes.

Before you start¶

A FLEX-8600 radio must be connected and active.
The whisper.cpp engine requires a model file; on first use, it will download the selected model automatically.
GPU acceleration requires a compatible NVIDIA GPU (with CUDA) or Apple Silicon Mac (with Metal) and sufficient VRAM. On hosts without a compatible GPU, the engine runs in CPU mode.

How it works¶

Open the Copy Assist panel via View > Copy Assist or press Ctrl+Shift+T.
Enable transcription by clicking the Enable / Disable toggle button. The engine state indicator will show "Downloading model" (first use) → "Loading model" → "Listening".
Monitor the transcript in real time as decoded text scrolls in the read-only Transcript field. Text is color-coded by confidence:
Green – high confidence
Yellow – medium confidence
Red – low confidence
Track pipeline health using the Backlog indicator — shows seconds of audio not yet transcribed. The indicator color escalates from amber to red as backlog grows.
Fine-tune performance by clicking the Settings button (gear icon) to open the modeless Copy Assist settings dialog, where you can adjust:
Model tier – choose tiny, base, small, or medium (larger = more accurate but slower and uses more VRAM/RAM)
Compute device – select GPU (CUDA/Metal) for faster inference with GPU or CPU for universal compatibility. GPU availability is probed asynchronously on startup; the engine uses the platform default until the probe completes (on macOS the first GPU query can take several seconds while the Metal shader library compiles).
Clear the buffer at any time by clicking the Clear button. This also resets the whisper decode context so the next transcription starts from a clean prompt.

Note: The Clear button and the transcript's internal clear both flush the carried decode context on the whisper backend, ensuring the fresh display starts with no stale prompt state.

What each control does¶

Control	Type	Default	Behavior
Enable / Disable	toggle button	Disabled	Starts or stops the whisper.cpp engine on the active slice's audio.
Transcript	read-only text field	—	Scrolling text display of decoded speech. Color-coded by confidence: green (high), yellow (medium), red (low). Includes a Clear button that empties the buffer and resets decode context.
Model tier	combo box	`tiny`	Selects whisper model size. Valid settings: `tiny`, `base`, `small`, `medium`.
Compute device	combo box	`GPU (CUDA/Metal)`	Selects inference device. Valid settings: `GPU` or `CPU`.
Backlog indicator	status indicator	0.0s	Shows seconds of audio in the pipeline not yet transcribed. Color escalates amber → red as backlog grows.
Settings button	push button	—	Opens the modeless Copy Assist settings dialog for model tier, compute device, and engine configuration.
Clear	push button	—	Clears the current transcript buffer and resets the whisper decode context.

Engine status indicator¶

The Copy Assist panel shows the current state of the speech-to-text engine:

State	Meaning
Idle	Engine is not running.
Downloading model	The selected model file is being downloaded on first use.
Loading model	The model file is being loaded into memory.
Listening	Engine is actively decoding audio from the active slice.
Error	The engine encountered a problem (e.g., model download failure, backend unavailable).

Tips¶

Start with the tiny model for lowest latency and minimal resource usage. Switch to base or small if accuracy is insufficient and your system can handle the load.
GPU mode is significantly faster but requires a compatible GPU. If you experience high backlog times, try switching to GPU (if available) or reducing model size.
An explicit compute-device choice is remembered and takes priority over the automatic GPU detection. If you later change hardware, clear the saved preference by selecting a different device or removing the setting via the configuration file.