EchoRadar v2

EchoRadar v2 is Windows research software that captures headphone playback audio and turns discrete 5.1/7.1 system-output audio into a 24-sector azimuth radar. It uses an original frequency-domain multichannel energy-vector implementation and a tactical-dark dashboard/HUD. It does not use ASUS code, assets, branding, or proprietary tuning, and it does not claim bit-for-bit compatibility with ASUS Sonic Radar.

The playback signal path preserves native channel roles:

Windows WASAPI loopback (stereo or WAVEFORMATEXTENSIBLE 5.1/7.1)
  -> 48 kHz analysis PCM, retaining the endpoint channel map
  +-> independent audio activity / headphone left-right energy monitor
  +-> 2048-sample Hann STFT / 480-sample hop
  |    -> per-bin multichannel energy vector
  |    -> 24 azimuth sectors + smoothing -> dashboard and HUD
  +-> LFE-free stereo downmix -> unchanged stereo-onset-v4 recognizer
       -> optional event-window radar peaks

Stereo headphones support capture, activity monitoring, and optional event recognition. Full azimuth requires discrete 5.1/7.1 channels; sample-rate conversion is allowed when the channel-role map is preserved. EchoRadar v2 never invents surround bearings from stereo. The frozen v1 implementation remains at branch codex/v1-maintenance and annotated tag echoradar-v1.0.0.

Features

• Native Windows 5.1 and 7.1 channel-role preservation; LFE is excluded from directional analysis and recognition downmix.
• Continuous, event, and combined radar modes.
• All, Footsteps, Gunshots, and editable 16-point Custom spectral emphasis curves. Preset names describe frequency emphasis only; they do not classify sound semantics.
• 24 sectors at 15-degree spacing, with 0 degrees front, 90 degrees right, 180 degrees rear, and 270 degrees left.
• Resumable surround setup, actionable inline runtime states, a single keyboard-navigable dashboard, and a click-through/editable per-display HUD.
• Settings schema 4 under %LOCALAPPDATA%\EchoRadar\v2 and JSONL schema 3 session records.
• Optional existing stereo-onset-v4 recognition packages run in every radar display mode. Audio activity and continuous surround radar need no model.

Build

Windows 10/11 x64, CMake 3.20+, a C++20 toolchain, and DirectX 11 are required for the application. ONNX Runtime CPU is required only when recognition is enabled. The first configure fetches pinned KissFFT, miniaudio, ImGui, and—when tests are enabled—GoogleTest.

cmake -S . -B build `
  -DECHORADAR_BUILD_APP=ON `
  -DECHORADAR_BUILD_AUDIO_MONITOR=ON `
  -DECHORADAR_BUILD_TESTS=ON `
  -DECHORADAR_ENABLE_ONNX=ON `
  -DONNXRUNTIME_ROOT="C:\path\to\Microsoft.ML.OnnxRuntime"
cmake --build build --config Release
ctest --test-dir build -C Release --output-on-failure

The application target is EchoRadarV2.exe.

Run

.\build\src\app\Release\EchoRadarV2.exe `
  --radar-mode continuous `
  --radar-preset all

Recognition is optional:

.\build\src\app\Release\EchoRadarV2.exe `
  --model models\recognition-candidate `
  --radar-mode combined `
  --radar-preset footsteps

When --model is omitted, EchoRadar looks for models\recognition-candidate and then the legacy-compatible models\v4-candidate package name.

Use --list-audio-outputs to print native formats and ordered channel roles, --audio-output-id <id> to pin an endpoint, --settings <json> to override the v2 settings file, and --no-overlay for a headless run.

Accuracy and limitations

EchoRadar reports azimuth only. Sector peaks are energy maxima, not proof of separate physical sources. Opposing sounds with energy in the same frequency bins can cancel in an energy vector. Room reflections, game mixing, endpoint processing, virtual-surround drivers, and incorrect Windows speaker masks can reduce accuracy. Discrete 5.1/7.1 channel roles are required for full azimuth. Stereo energy balance cannot distinguish front from rear.

The automated suite covers channel mapping, buffering/downmix, radar bearings, interpolation, presets, smoothing, cancellation, state transfer, settings, and HUD bounds. Real endpoint and visual/DPI results are tracked as release gates in docs/evaluation.md; unmeasured gates are not reported as passes.

Draft release notes and the publication checklist are in docs/release-v2.0.0-research.md.

License and patent notice

EchoRadar v2 is licensed under the Apache License 2.0. Apache-2.0 grants patent rights from contributors only; it does not grant rights under third-party patents. See NOTICE and algorithm provenance, including the notice for US 9,232,337 B2.

Local sounds/, models/, recordings/, generated corpora, runs, caches, and checkpoints are ignored and must not be redistributed without appropriate rights.

The headphone investigation, implemented changes, remaining full-spatial work, and real-device checklist are in audio/headphone refactor.