Denoise vs DPDFNet

Denoise is an audio cleanup app for Android and iOS that uses an on-device neural network to reduce wind, traffic, hiss, hum, and crowd noise in voice memos, interviews, and videos. Processing runs locally, including in airplane mode, and a one-minute clip typically takes 5–10 seconds. Users can import recordings from Voice Memos, Photos, Files, or an app share sheet; supported inputs include MP4, M4A, MP3, and WAV. The free plan lets users import, process, and hear a before-and-after preview, but saving and export require Pro. Pro can export M4A, MP3, or WAV, replace audio inside a video, restore detail in clipped, muffled, or distant recordings, and separate vocals from instrumental tracks on-device. Audio tools include a 5-band EQ with presets and automatic LUFS normalization. There is no file-length limit, including for recordings up to a three-hour lecture. Routes Software says recordings are neither uploaded nor stored, and the app has no user accounts. Google Firebase collects anonymous usage analytics and crash and diagnostic reports. Pro costs $2.99 a month or $14.99, paid once for lifetime access.

DPDFNet is a free, open-source model for reducing noise in single-channel speech from recordings, live streams, and edge-device applications. It supports offline processing and streaming, and extends DeepFilterNet2 with dual-path encoder blocks and an always-on fine-tuning approach intended to reduce over-attenuation. Pretrained variants support 8, 16, and 48 kHz audio. The repository includes PyTorch code and ONNX and TFLite inference models, with command-line, Python, and stateful streaming interfaces. Streaming buffers about 20 ms before the first enhanced output, then adds about 10 ms of delay for later blocks. A hosted Gradio demo and audio examples are linked from the project page. Listed community integrations include sherpa-onnx, OBS Studio, VST3 and LADSPA plugins, Core ML ports, and WebAssembly. The project page reports real-time DPDFNet-4 deployment on Ceva-NeuPro-Nano edge NPUs in a benchmark setup using int8 weights and int16 activations. The software is under Apache License 2.0. Model files are not bundled in the source repository; it provides commands to download checkpoints and ONNX or TFLite models from Hugging Face.