Select your microphone from the device selection menu at the top.
RNNoise requires a consistent buffer size. If your ASIO buffer is too high (e.g., 1024 samples), the recurrent neural network loses context. Additionally, RNNoise is trained on standard male/female vocals; whispering or extremely high-pitched singing may be misclassified. Fix: Reduce buffer size to 128 or 256 samples. For singing, use a traditional noise gate instead. librnnoise-vst.dll
At its core, the filename itself is a semantic roadmap. The prefix lib (standard for "library") indicates a collection of reusable functions. rnnoise is the true identifier: it stands for . This is an open-source project conceived by Jean-Marc Valin, a renowned audio engineer at Mozilla (and co-creator of the Opus codec). Unlike traditional noise gates or spectral subtraction algorithms that work on static thresholds, RNNoise uses a deep learning model trained on thousands of hours of clean and noisy speech. The suffix -vst is the most critical qualifier. VST (Virtual Studio Technology) is a software interface standard developed by Steinberg, allowing third-party audio effects (reverb, compression, equalization) to run inside Digital Audio Workstations (DAWs) like Ableton Live, FL Studio, Reaper, or Audacity. Finally, .dll signifies that on Windows, this is a dynamically linked library—a chunk of executable code that loads only when needed. Select your microphone from the device selection menu