Remove noise from audio streams

Use This When

• Preprocessing audio before speech recognition to improve transcription accuracy
• Building voice interfaces where background noise degrades user experience
• Recording or streaming in noisy environments like factories, vehicles, or outdoors
• Enhancing audio quality for downstream processing or human consumption

What It Does

• Applies MPSENet deep learning model to suppress background noise while preserving speech
• Resamples audio to 16kHz for model processing then outputs denoised frames
• Supports model hot-swapping without restarting pipeline via mutable model config
• Processes audio frame-by-frame maintaining stream continuity

Works Best With

• Audio inputs → this component → VAD, transcription, or voice activity detection
• Voice UI pipelines where clean audio improves downstream ASR accuracy
• Recording workflows needing noise suppression before storage or transmission

Caveats

• Resamples all audio to 16kHz regardless of input rate; output is always 16kHz
• Aggressive noise suppression can introduce artifacts or remove weak speech
• Model trained on specific noise types; performance varies with novel noise profiles
• GPU strongly recommended; CPU inference adds significant latency