Jivetalking 0.5.1

Highlights ✨

Validated against the quality reference across the full corpus: equal or better loudness and dynamics, with a faster pipeline.

Audio quality

• Speech detection now catches conversational speakers (gate lowered to 10s) — every presenter elects a speech profile, so filters adapt off speech-only measurements.
• De-esser rebuilt on sibilant-vs-body band excess; it arms on genuine sibilance instead of a discredited centroid model.
• Noise reduction swaps compand for afftdn — cleaner under-speech noise with less floor pumping.
• Room-tone election picks a representative region by proximity rather than chasing the quietest sample.
• Ceiling margin and crest handling now derive from the validation corpus.

Speed

• Faster anlmdn path and tightened adeclick cut Pass 2/4 runtime at metric-parity quality.

Reporting accuracy

• Reports show actual peak reduction (not zero dB), flag loudnorm dynamic-mode fallback, and use corrected noise-floor labelling.

Internals

• Removed dead adaptive "theatre" across the gate, LA-2A, de-esser, and HP/LP — adaptation that never exercised its range is gone.
• ffmpeg-statigo updated to v8.1.1.2; dependency bumps.

Changes since 0.5.0

• fix(processor): derive ceiling margin from validation corpus (9e72033)
• refactor(processor,audio): clarify callback semantics, frame aliasing, and cgo cleanup (ec38f20)
• chore(deps): bump nixpkgs from 331800d to a799d3e (6e7bce7)
• deps(deps): bump github.com/mattn/go-runewidth in the gomod group (a0e4a46)
• refactor(processor): delete theatre from DS201 gate adaptive surface (e7736a2)
• fix(logging): detect and report loudnorm dynamic fallback in mode headline (e372376)
• fix(logging): report actual peak reduction instead of zero dB (e44fa59)
• refactor(processor): simplify LA-2A to single threshold adaptation (c4729f6)
• fix(processor): replace de-esser centroid model with band-excess engagement (31019cb)
• refactor(processor): remove adaptive DS201 highpass and LP formatter theatre (46ec72c)
• fix(logging,filters): correct noise-floor label and gain-normalisation format, extract afftdn helper (6585473)
• fix(processor): replace fixed crest cap with population-relative outlier gate (583e23e)
• refactor(processor): remove dead fallback in room-tone candidate election (54b1f17)
• fix(processor): elect representative room tone by proximity instead of amplitude monotonic (6174cfa)
• fix(processor): lower speech detection gate to 10s for conversational speakers (098ef6c)
• refactor(processor): replace compand with afftdn in noise-reduction block (3fcc67f)
• refactor(processor): make DS201 lowpass unconditional 20.5 kHz band-limit and tighten adeclick threshold to 1.7 (65f07d1)
• chore: update ffmpeg-statigo to v8.1.1.2 (d62d42c)

Installation

Download the appropriate binary for your platform below, make it executable, and move it to your PATH:

# Linux (amd64)
chmod +x jivetalking-linux-amd64
sudo mv jivetalking-linux-amd64 /usr/local/bin/jivetalking

# macOS (Apple Silicon)
chmod +x jivetalking-darwin-arm64
sudo mv jivetalking-darwin-arm64 /usr/local/bin/jivetalking

Checksums

SHA256 checksums are provided below for verification.