Skip to content

Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more

3.3 KB
Raw Blame

Fluid WebRTC Audio Processing package

This local Swift package contains FluidVoice's pinned WebRTC Audio Processing Module/AEC3 bridge. It is built from Google WebRTC revision d9bd07ba5f614156021df666ba4052ac73cf7953 and statically linked for macOS 15 on arm64 and x86_64.

The product build consumes the checked-in static XCFramework and therefore needs neither a WebRTC checkout nor network access. The XCFramework was built locally from the pinned source; it is not a downloaded framework. UpstreamSources contains every upstream and generated compile input used by the archive members that the bridge executable actually linked. Metadata records the full recursive GN target closure, selected archive-object dependency graph, exact checkout revisions, build settings, and SHA-256 hashes. BridgeSources contains the C ABI implementation and its standalone probe.

The full GN closure is intentionally larger than the final static archive. WebRTC's APM graph still contains build-time Perfetto/protobuf generator edges even with rtc_enable_protobuf=false. The link-map-derived archive contains no protobuf or Perfetto object, no concrete AEC dump writer, and uses WebRTC's null AEC dump factory. All public app-facing declarations are C ABI; WebRTC C++ headers are not exposed by the package product.

The static archive intentionally has no root module.modulemap: callers import the source FluidAEC3Bridge module. A binary module map would collide with other static XCFrameworks (such as Zeppelin) when Xcode copies their headers into the shared build include directory.

Run Scripts/verify.sh to verify hashes, architectures, symbols, revision/configuration strings, and a clean SwiftPM link. Scripts/refresh.sh documents and automates the reproducible rebuild from an exact gclient checkout plus depot_tools. The refresh script rejects any revision mismatch. Scripts/benchmark.sh builds an optimized, current-architecture harness and processes a 30-minute deterministic synthetic stream after warm-up. It enforces p99 below 5 ms, every call below 10 ms, bounded RSS variation, and no monotonic RSS growth.

Licenses and required attribution for WebRTC and linked third-party source are in Notices.

Meeting speech protection

The v2 bridge keeps the pinned WebRTC revision and default echo-only suppression. It enters local-speech protection sooner and relaxes the low/high-band residual-echo mask only during detected local speech. It adds no capture delay, gain, noise suppression, extra model, or realtime statistics lock. This improves quiet speech under loud speaker playback; some residual echo during overlapping speech remains possible. It does not guarantee separation of every speaker mixture.

BridgeSources/rebuild_bridge.sh recompiles the bridge for both architectures using the vendored headers and system libc++, replacing only fluid_aec3_bridge.o in each verified archive slice. It preserves all pinned upstream objects, runs the contract, strong-voice and weak-voice overlap probes, and updates the two integrity manifests. It refuses modified upstream inputs. This is a bridge-only rebuild, not a new GN build of upstream WebRTC; the manifest records that distinction. The weak-voice regression correlates output with an independent known voice so leaked playback cannot count as preserved speech. XCTest repeats that check through the app wrapper.