Talon
bad9c7533a
The two-sided NR plumbing (RemoteStream::recv_ns + the per-listener vc_set_remote_stream noise_reduction toggle) was wired but inert: ApmProcessor::create() returned a no-op passthrough, because the originally-planned webrtc-audio-processing has no working Windows/macOS build. Drop in RNNoise as the real backend behind the same ApmProcessor interface, lighting up both NR paths. - Vendor RNNoise (BSD-3 + CC0) at third_party/rnnoise/ — the vcpkg port is !windows !arm, so it can't cover our primary targets. Shrunk int8 model (78MB -> 11.7MB via upstream scripts/shrink_model.sh), built as a standalone C static lib with no RTCD (portable scalar path on x86, auto-NEON on arm64) under -DDISABLE_DEBUG_FLOAT. Model is baked in (rnnoise_create(NULL)); no runtime file. - New RnnoiseProcessor (core/src/audio/apm_processor.cpp) selected by ApmProcessor::create() when VOICECAT_HAS_NS. Mono/48kHz/480-sample; our clock is fixed 48kHz and Opus frame sizes are multiples of 480, so no resampling. RT-safe: allocates at construction, lock-free in the capture/playback callbacks. - Receive-side: lit up via the factory; gated to mono streams (a stereo stream is a screen-audio share, not voice). - Send-side (new): vc_set_input_noise_reduction(client, enable) ABI + vc_client::mic_ns_, run before input gain/VAD in on_capture_frame. A stereo mic is downmixed to mono ONLY when NR is on — with NR off a stereo mic keeps full stereo (never collapse mic quality unasked). - Enable C as a project language for the vendored lib. - New noise_suppression test: white noise through ApmProcessor::create() drops ~99.9% RMS. ctest --preset dev green, 28/28. windows-client DLL builds clean with vc_set_input_noise_reduction exported, system-only deps. - Docs synced: voice.md §10, tech-stack.md §1/§5, third_party/README.md, vcpkg.json note, PROGRESS.md, CLAUDE.md. Client on/off UI toggles (Windows/macOS/iOS) are the remaining follow-up. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
third_party/ — vendored dependencies
Dependencies that are not consumed through vcpkg live here, copied verbatim into the tree.
Everything here is permissively licensed (no GPL/LGPL) per the house rule in
docs/tech-stack.md §5.
rnnoise/
Real-time speech noise suppression (the DSP backend behind ApmProcessor —
see docs/voice.md §10–11). Used by the receive-side per-stream NR
(RemoteStream::recv_ns) and the send-side mic NR (vc_client::mic_ns_).
- Upstream: https://github.com/xiph/rnnoise
- Vendored at commit:
70f1d256acd4b34a572f999a05c87bf00b67730d - License: BSD-3-Clause (code, see
rnnoise/COPYING) + CC0-1.0 (model weights). - Why vendored, not vcpkg: the vcpkg
rnnoiseport is marked!windows !arm, i.e. unavailable on our primary targets (Windows MinGW, Apple Silicon, iOS). RNNoise is small, self-contained C99 with no dependencies, so we vendor it directly.
What was copied / changed
- Only the library sources + headers (
src/*.c,src/*.h,src/x86/*.h,include/). The training/feature-dump tools (dump_features.c,write_weights.c, thesrc/x86/*.cRTCD kernels), build scaffolding (autotools, Meson) andtorch/training/dirs are omitted. src/rnnoise_data.cis the shrunk model: upstream'sscripts/shrink_model.shstrips the#ifndef DISABLE_DEBUG_FLOATfloat-weight duplicates, taking the default model from ~78 MB to ~11.7 MB. We build with-DDISABLE_DEBUG_FLOATso only the int8-quantized weights are used — this is exactly how upstream's default (non-debug) build behaves. The model is the built-in default loaded byrnnoise_create(NULL); there is no runtime model file.
Build
Built as a standalone static lib rnnoise in core/CMakeLists.txt
(no RTCD; portable scalar path on x86, NEON on arm64), and linked into libvoicecat which
defines VOICECAT_HAS_NS. To refresh the model, re-run upstream autogen.sh/download_model.sh
scripts/shrink_model.shand re-copysrc/rnnoise_data.{c,h}.