Signal Flow

Everything RFWhisper does, in order, from antenna (or rig audio) to your decoder.

Stage-by-stage detail

1. Audio capture

• Backends: PortAudio (Linux/mac), WASAPI (Windows), CoreAudio (mac native), ALSA (Linux native).
• Frame sizes: 10–30 ms native (160–480 samples @ 16 kHz; 480–1440 @ 48 kHz). Default 10 ms at 48 kHz for DFN3.
• Sample format: float32 internally. Device-side conversion via the backend.
• Thread: realtime priority where OS allows.

2. SPSC rings

Single-producer single-consumer lock-free ring between capture and processing, and between processing and output. Tuned to ~4 frames of headroom; larger buffers trade latency for safety against device jitter.

3. Pre-DSP (rfwhisper/dsp/)

In order:

1. De-emphasis (if input is NBFM) — 50 µs / 75 µs curves.
2. Polyphase resample to model native rate (48 kHz for DFN3 / RNNoise-ham). Uses liquid_msresamp. Never scipy.signal.resample in the hot path.
3. Windowing (Hann by default; sqrt-Hann when the output chain also windows). VOLK-accelerated.
4. STFT via liquid-dsp FFT plans, plans reused across frames.
5. Feature prep — model-specific. DFN3 ingests spectrogram mag+phase; RNNoise uses its custom 42-dim bark features (handled inside the ONNX graph).
6. (v0.3) Adaptive narrowband notch — classical LMS / gated IIR that removes stable carriers before the NN. Off by default for CW/FT8 profiles.

All preallocated. Zero Python in the per-frame critical path (we use C extensions / numpy views with pre-cast buffers).

4. ONNX inference (rfwhisper/models/)

• Execution provider chosen by rank: CoreML → DirectML → CUDA → XNNPACK → CPU. Override with RFWHISPER_PROVIDER=....
• I/O tensors allocated once; bound via ONNX Runtime's IO binding API.
• intra_op_num_threads tuned per-target (see Performance / Embedded if you're reading in the repo).
• GIL released around OrtRun.

5. Post-DSP

1. iSTFT + overlap-add with matched window. Overlap-add windows must sum to 1.0 — verified in unit tests.
2. Gain match — output RMS matched to input RMS on a slow AGC to avoid perceived level jumps when toggling bypass.
3. Soft limiter — catches occasional NN overshoots, never touches normal levels.

6. Output ring + device

Symmetric to the input side. Drains to the virtual cable or speaker.

Per-profile variations

Mode profiles (YAML in rfwhisper/profiles/) tweak:

• NN aggressiveness (blend factor 0.0–1.0)
• Notch enable + bandwidth (v0.3)
• De-emphasis on/off
• Attack / release envelopes for the soft limiter
• Frame size (longer windows for FT8, shorter for CW transient preservation)

Profile	NN aggression	Notch	Frame	Rationale
ssb	0.9	off	10 ms	Speech intelligibility
cw	0.6	off	5 ms	Transient preservation is paramount
ft8	0.7	off	20 ms	Decoder tolerates latency; be gentler
ft4	0.8	off	10 ms	—
fm-narrow	0.9	optional	10 ms	De-emphasis first
vhf-ssb	0.85	optional	10 ms	Weak-signal-friendly

Latency measurements per stage live in Latency Budget.