Realtime zero-shot voice conversion. One short reference clip of the target voice (5 s+) is enough — no retraining. Runs on CPU with a hand-written C engine that streams through a HiFi-GAN generator and keeps RTF < 1 on modest hardware.

The desktop app is an Electron shell over libvc.dll.

• C engine (cpp/) — production backend (OpenBLAS + RNNoise + miniaudio, streaming decoder, pool-based parallel resblocks).
• Electron UI (electron/) — device picker, meters, AGC slider, VAD gate.

Prebuilt Windows zip on the Releases page. Unzip and run Lilac.exe — libvc.dll, weights.bin, and tsu_10s.wav sit alongside the executable and must stay in the same folder.

Requires MSYS2 (mingw-w64-x86_64-toolchain, mingw-w64-x86_64-gcc-fortran) plus a prebuilt OpenBLAS tree at cpp/openblas/ and cpp/weights.bin.

cd cpp
mingw32-make libvc.dll

Output cpp/libvc.dll has OpenBLAS, gfortran, and RNNoise statically linked — no sibling DLLs need to ship.

cd electron
npm install
npm start

cd electron
npm run dist

Output lands in electron/dist/win-unpacked/ with Lilac.exe, libvc.dll, weights.bin, and tsu_10s.wav at the top level; zip the folder to distribute.

• 48 kHz device I/O via miniaudio.
• RNNoise runs on the capture thread (10 ms frames) for VAD; raw audio feeds the VC path, with an 80 ms VAD hangover to preserve speech tails.
• AGC (target −20 dBFS default) between VAD and VC.
• 48 ↔ 22.05 kHz stateful resampling.
• VC worker runs the streaming HiFi-GAN generator; output has a 600 ms silence-gate hangover so the engine's internal lookahead can flush before bias is blanked.
• Two-thread pipeline (rnn + vc) connected by SPSC ring buffers.

See Optimization.md for the full list of perf measures.

Source: https://github.com/kdrkdrkdr/lilac