You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
GTCRN seems like the better first target for this repo: it is small enough to fit ExecuTorch's mobile footprint expectations, and its streaming formulation matches the way RN audio pipelines usually consume frames.
A practical path would be:
Start with GTCRN ONNX and run it through ExecuTorch's export/verification path before writing native bindings.
Keep the model behind an explicit audio preprocessor API rather than baking it into STT or call code.
Validate on 10/20 ms frame budgets, not just offline clips: mobile CPU scheduling and background thermal limits are usually the limiting factor.
Keep DeepFilterNet3 as a later quality reference, but I would not make it the default mobile model unless the binary/memory/CPU budget is clearly acceptable.
The main question is whether the project wants a general-purpose audio enhancement module or a narrow "clean input for STT" module. The first is more reusable; the second is easier to keep small and predictable.
Activity
mPiha commentedon Nov 10, 2026
GTCRN seems like the better first target for this repo: it is small enough to fit ExecuTorch's mobile footprint expectations, and its streaming formulation matches the way RN audio pipelines usually consume frames.
A practical path would be:
The main question is whether the project wants a general-purpose audio enhancement module or a narrow "clean input for STT" module. The first is more reusable; the second is easier to keep small and predictable.