StreamLite 2.3.0

• Low VRAM (Windows): a new switch on the Microphone page, under noise removal (also on the Camera page), for graphics
  cards with little video memory, such as a 4 GB RTX 4060. DeepFilterNet on the GPU and the AI camera then use small
  versions of their models. Everything together stays under 250 MB of video memory, down from about 430 MB, at the
  same speed:

  	Normal	Low VRAM
  DeepFilterNet 3 (GPU)	~150 MB	~60 MB
  + AI camera (1080p)	~430 MB	~210 MB

  Measured on an NVIDIA card with Windows' GPU counters (dedicated video memory). Windows also lists about 60 MB of
  "shared GPU memory"; that is ordinary RAM, not video memory.

  • DeepFilterNet keeps its recurrent layers in 16-bit floats. Your voice comes out within 50 dB of the full model.
  • The AI camera runs its effects in 16-bit floats with fewer full-size images in flight. The picture matches the
    normal model (46-58 dB PSNR). Capture is 1080p while Low VRAM is on (4K capture needs the normal model).
  • DirectML runs these models without the driver's "metacommands", which reserve a lot of memory for no gain here.
  • The CPU engines are unchanged. If the graphics card can't run DeepFilterNet, it still moves to the CPU.

Credits

Low VRAM: onnxconverter-common (half-precision model files) and ONNX Runtime's model optimizer build the new deepfilternet3_low.onnx and camfx_1080_low.onnx, which are made from DeepFilterNet 3, Robust Video Matting and UltraFace.

AI models and measured data

• DeepFilterNet 3 - AI noise removal (studio-quality model). Hendrik Schröter, Tobias Rosenkranz, Alberto N. Escalante-B., Andreas Maier. MIT or Apache-2.0. Model / data.
• torchDF - Streaming ONNX export of DeepFilterNet 3 (the model file StreamLite runs). grazder. MIT or Apache-2.0.
• RNNoise - Light AI noise removal (C# port of v0.1.1 and its trained model). Jean-Marc Valin, Mozilla, Xiph.Org Foundation. BSD-3-Clause. Model / data.
• SADIE II HRTF database - 3D audio: measured head responses (KU100 dummy head, KEMAR). University of York - C. Armstrong, L. Thresh, D. Murphy, G. Kearney. Apache-2.0.
• Robust Video Matting (Windows) - AI camera: cuts you out of the room (blur, green screen, picture, video). Shanchuan Lin, Linjie Yang, Imran Saleemi, Soumyadip Sengupta (ByteDance). GPL-3.0. Model / data.
• Ultra-Light-Fast-Generic-Face-Detector-1MB (UltraFace) (Windows) - AI camera: face tracking. Linzaer. MIT. Model / data.
• SoundGuys headphone measurements - Built-in correction for the Sennheiser MOMENTUM 5 (B&K 5128 measurement). SoundGuys. Measurement used with thanks.

Libraries and runtimes

• ONNX Runtime - Runs the AI models (and, from 2.3.0, optimizes the Low VRAM model files). Microsoft. MIT.
• DirectML (Windows) - Runs the AI models on any DirectX 12 graphics card. Microsoft. Microsoft Software License Terms (redistributable).
• .NET runtime (Linux) - Runs StreamLite for Linux (included). .NET Foundation and contributors. MIT.
• Avalonia UI (Linux) - The Linux interface. AvaloniaUI OÜ and contributors. MIT.
• SkiaSharp / HarfBuzzSharp (Linux) - Drawing and text for the Linux interface (Skia, HarfBuzz). Microsoft / .NET Foundation; Google (Skia); the HarfBuzz authors. MIT; BSD-3-Clause (Skia); MIT (HarfBuzz).
• Inter (Linux) - Interface font. Rasmus Andersson. SIL Open Font License 1.1.
• PipeWire / PulseAudio (Linux) - Linux audio (your system's libpulse and pactl). Wim Taymans and the PipeWire contributors; the PulseAudio developers. MIT / LGPL-2.1.

Code, ideas and research

• EarTrumpet (Windows) - How per-app audio routing works on Windows. File-New-Project. MIT.
• Impulcifer - HeSuVi 14-channel HRIR track order. Jaakko Pasanen. MIT.
• VCamSample (Windows) - Structure of the Windows 11 virtual camera. Simon Mourier. MIT.
• Guided Image Filtering (Windows) - AI camera: denoise, skin smoothing and crisp cut-out edges. Kaiming He, Jian Sun, Xiaoou Tang. Published research.

Downloaded when you ask for them

• AutoEq - Headphone corrections for thousands of headphones (Find my headphones). Jaakko Pasanen and the measurement contributors (oratory1990, crinacle, Rtings, Innerfidelity...). MIT. Model / data.
• VB-CABLE and Hi-Fi Cable (Windows) - The virtual mic cable and the 3D game audio device. VB-Audio Software (Vincent Burel). Donationware - please support VB-Audio.
• Equalizer APO (Windows) - System-wide headphone EQ. Jonas Thedering. GPL-2.0.

Built with

• WiX Toolset (Windows) - The Windows installer (MSI). .NET Foundation and contributors. MS-RL.
• Zig (Windows) - Builds the virtual camera's native DLL. Andrew Kelley and the Zig contributors. MIT.
• appimagetool (Linux) - The Linux AppImage. Simon Peter and the AppImage contributors. MIT.
• ONNX - Builds and adapts the model files (tools/make_camfx.py). ONNX Project Contributors (LF AI & Data). Apache-2.0.
• NumPy - Model-building scripts. NumPy Developers. BSD-3-Clause.
• onnxconverter-common - Half-precision Low VRAM model files. Microsoft. MIT. New in 2.3.0.

---

Which file do I need?

Your system	File
Windows 10/11	StreamLite-*-x64.msi. If SmartScreen warns you, click More info → Run anyway.
Windows, no install	StreamLite-*-win-x64-portable.zip
Ubuntu / Debian / Mint / Pop!_OS	streamlite_*_amd64.deb, then run sudo apt install ./streamlite_*_amd64.deb
Fedora / openSUSE	streamlite-*.x86_64.rpm
Arch / Manjaro / EndeavourOS	PKGBUILD, then run makepkg -si
Any Linux	StreamLite-*-x86_64.AppImage, or the linux-x64.tar.gz with ./install.sh

SHA256SUMS.txt lists the checksums of every file. Already have StreamLite? It updates itself when it starts.
New here? Read the manual. Something wrong? Use the
bug report.