These features will help manage VRAM. Very useful if you're gaming or similar.
Unload models when idle
Local models can now leave memory when you are not using them.
{
"model_idle_unload_minutes": 10
}After ten idle minutes, the model unloads to free RAM or VRAM. The next recording loads it again while you speak. Text arrives once the model is ready.
The default is 0, which keeps the model loaded.
If loading fails, the audio is kept and the notification gives you a command to transcribe it. Cancelled recordings are discarded. If saving fails too, the notification says so.
File transcription also wakes an idle-unloaded model. Long-form recording starts the reload and asks you to try again once it is ready.
Unloading by hand still holds until hyprwhspr model reload, including across suspend and resume.
Parakeet VAD batch size
The number of VAD segments processed together is now configurable:
{
"onnx_asr_vad_batch_size": 1
}The default remains one segment at a time. Higher values can speed up long GPU recordings at the cost of more memory.
The service log now reports the execution provider actually in use, including CPU fallback. Transcription timing includes processing the VAD segments.
Fixes
- Continuous recording: An empty or filtered silent tail after a successful paste no longer plays an error beep.
- Chunk ordering: Stopping waits for an in-flight continuous chunk before processing the remaining audio.
- Recording feedback: Filtered hallucinations show their result once.
- CLI recording: Start commands wait for confirmation before reporting that recording started. Long-form recording counts toward recording status and actions.
- Model reload: The command now reaches models unloaded by the idle timer and reports when the model is already loaded.
- Model recovery: Resume reinitialization is serialized with transcription and model loading, with state checked again before proceeding.
Thanks to @robstiles! 🙏
Full Changelog: v1.53.1...v1.54.0