Loudnorm per transmission

[TheGreatCodeholio]

Per-transmission audio post-processing pipeline

Summary

Two related changes that together rework how call audio is filtered and normalized:

1. Per-transmission loudnorm + final limiter — the old two-pass linear loudnorm applied a single static gain to the entire concatenated call, so a quiet capture and a loud capture in the same call stayed quiet and loud
   relative to each other. Now each transmission is independently cleaned, normalized, and limited before being concatenated, so speech sits at the same target loudness across every transmission in a call.

2. Analog bandpass moves from GNURadio into ffmpeg post-processing — the hardcoded 300/3000 Hz FIR bandpass that lived inside analog_recorder is now audio_postprocess.highpass_hz / lowpass_hz (defaulting to those
   same values). This makes the bandpass user-tunable, drops two realtime FIR filters from every analog recorder, and means the original transmission WAVs are now real post-demod / post-deemph discriminator audio — so
   outputRawAudio and transmissionArchive finally deliver what they advertise.

Along with that, audio_postprocess.enabled becomes the master switch for the whole pipeline. audio_postprocess.enabled=false skips cleanup, loudnorm, and the limiter; the call's main file becomes a stream-copy concat of the raw
transmission WAVs (the same shape outputRawAudio produces as a side file).

Why this matters

Measured on a real multi-transmission call (3 transmissions, ~20 s):

	Before	After
Integrated loudness	-17.1 LUFS	-15.8 LUFS
Loudness range (LRA)	25.0 LU	3.3 LU
True peak	+0.3 dBFS (clipping)	-0.5 dBFS (safe)
Spread between quietest & loudest tx	~28 dB	~5 dB

LRA dropped 7.6×, the peak no longer clips, and listeners no longer hear "whisper then shout" within a single call. Quiet transmissions get pulled up to the loudness target; loud ones get tamed.

The two-pass loudnorm pipeline also cost ~2× the CPU because it ran an analysis pass and then a render pass. The new pipeline is a single ffmpeg invocation, equivalent in cost to the old single-pass mode and roughly half the
cost of the old default (two-pass).

What changed

Audio post-processing pipeline

• New per-transmission filter graph: each input gets cleanup → loudnorm → limiter, then all per-tx chains feed concat. Single ffmpeg invocation via -filter_complex.
• Removed the two-pass analysis path entirely (analyze_loudnorm_from_concat, LoudnormMeasured, build_loudnorm_analysis_filter, build_loudnorm_render_filter, and the run_process_capture_combined_output plumbing they
  used).
• Removed loudnorm_two_pass config field. It was also a latent bug — it was never copied from the system into call_info, so the user's config never took effect anyway.
• Added final_limiter config (default true). Ceiling derived from loudnorm_tp + 0.5 dB, attack 1 ms, release 50 ms.
• New voice-tuned defaults: loudnorm_tp -0.1 → -1.5, loudnorm_lra 11.0 → 7.0.

enabled as master switch

• audio_postprocess.enabled=false now skips the entire pipeline (cleanup + loudnorm + limiter), not just cleanup.
• loudnorm and final_limiter only take effect when enabled=true.
• Config parser logs a WARN at startup if a user has explicitly set enabled=false in their config so the semantic change can't surprise anyone silently. Migration note added to CONFIGURE.md.

Analog GR chain

• Deleted the high_f / low_f FIR blocks, their taps, and their connections from analog_recorder.{h,cc}. Chain is now decim_audio → squelch_two → levels → converter → wav_sink.
• decoder_sink already tapped at decim_audio (pre-bandpass), so CTCSS/MDC/etc. decoder behavior is unchanged.
• Defaults flipped so existing analog systems sound essentially the same out of the box: audio_postprocess.enabled=true, highpass_hz=300, lowpass_hz=3000.

transmissionArchive

• When audio_postprocess.enabled=true and transmissionArchive=true, per-tx archive files are written into the capture directory already processed (cleanup + loudnorm + limiter applied). They're produced in the same ffmpeg invocation as
  the main render via asplit on each per-tx pad, so no extra encode pass.
• When enabled=false, archived per-tx files are copies of the recorder's raw WAVs (existing behavior — and they're now meaningfully more raw for analog systems).
• Render only bothers writing archives when they'll actually be kept (gated on audio_archive || archive_files_on_failure), avoiding wasted disk writes.

Custom ffmpeg_filter

• Still works the same way — replaces the structured cleanup when enabled=true.
• Now applies per transmission rather than once on the concatenated stream. For chains that contain loudnorm, the user's loudnorm runs per-tx (and the built-in loudnorm is skipped, same rule as before).

Behavior changes / migration

Most users will see no behavior change beyond uniform speech loudness. The cases that do change:

Scenario	Before	After
Default config	GR-FIR 300/3000 bandpass, single per-call loudnorm	ffmpeg biquad 300/3000 bandpass, per-tx loudnorm + limiter (uniform speech)
audio_postprocess.enabled=false set explicitly	Cleanup skipped, loudnorm still ran	Cleanup, loudnorm, and limiter all skipped — main file is raw concat. Migration WARN logged at startup.
transmissionArchive=true (with audio_postprocess.enabled=true)	Raw recorder WAVs copied to capture dir	Per-tx files in capture dir are fully processed and playable standalone
Custom ffmpeg_filter set	Applied once to concatenated stream	Applied per transmission — for cleanup chains this is correct/better; for chains containing loudnorm, the loudnorm now runs per-tx (improves uniformity)
outputRawAudio=true	.raw.wav was bandpass-filtered	.raw.wav is true for analog that makes it post-demod / post-deemph discriminator audio for digital post vocoder audio