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
FluidAudio Nemotron compatibility patch
Base FluidAudio revision: 3fd63887eef1dc25edea8263ce4b44aa854d898b.
fluidaudio-nemotron-compatibility.patch preserves the independent dependency work. The current
development workspace points the app at that maintained local checkout; this must become a
published immutable fork revision before merge. Apply the patch at the base revision with
git apply --check followed by git apply. Do not patch Xcode DerivedData.
The current working checkout is .local-dependencies/FluidAudio, branch codex/nemotron-diarization;
that workspace is ignored, and this patch includes all nine changed/new source and test files.
Implemented: configurable speaker count, explicit Nemotron tensor geometry, strict shape/dtype and length validation, Float16 output decoding, caller-controlled compute configuration, local compiled model loading, cancellation checks around loading, and overflow-checked dimension validation. The Nemotron factory preserves the explicitly supplied update cadence, including 300, rather than silently applying the older Sortformer clamp to 340. Cache scoring parameters remain explicit. The host mel ABI is frozen per model family: Nemotron uses the checked-in NeMo settings and frame-count semantics, while legacy Sortformer retains its existing framing behavior. Complete-file Nemotron processing emits the final core tail instead of leaving it as right-context-only audio. Streaming state rejects inconsistent buffers and invalid reported embedding lengths.
Verified: the 67-test Sortformer/legacy integration selection passed, and the 10-test Nemotron
compatibility suite passed with the actual local CoreML artifact. The real-model suite loads the
model and runs one complete fixed window plus its final core tail end-to-end. It is enabled with
NEMOTRON_DIARIZATION_MODEL_PATH pointing to an .mlpackage or .mlmodelc directory. No model weights
are included in the patch or the DMG.
This establishes the Swift runtime contract and end-to-end execution. Model quality and performance measurements are intentionally out of scope because they were completed separately. Once the dependency changes are committed to the maintained fork, update the app's Xcode and Swift package pins together.