From 464edbc4fd658728f06d3a7d98340aa28ff3e3ec Mon Sep 17 00:00:00 2001 From: historia Date: Sat, 12 Sep 2026 15:25:06 -0400 Subject: feat: slop web gui --- config.toml | 24 +++++++++++++++++------- 1 file changed, 17 insertions(+), 7 deletions(-) (limited to 'config.toml') diff --git a/config.toml b/config.toml index 75e0e3b..a6777e3 100644 --- a/config.toml +++ b/config.toml @@ -5,9 +5,9 @@ # with --profile radio. Remove settings you want a different profile to supply. profile = "narrator" # natural, narrator, radio, cleanup-only -denoiser = "clearvoice" # clearvoice (MossFormer2_SE_48K, best attack preservation), deepfilter; fft is explicit non-AI fallback; none +denoiser = "clearvoice" # MossFormer2_SE_48K; compare with deepfilter on your voice; fft or none also available device = "auto" # auto, cpu, cuda; explicit cuda fails if unavailable -denoise_strength = 1.0 # 0..1; AI dry/wet blend; zero skips denoising and AI setup +denoise_strength = 0.85 # 0..1; a little dry signal can preserve consonants, but also retains noise fft_reduction_db = 8.0 # 0.01..30; FFT reduction is scaled by denoise_strength highpass = true @@ -26,11 +26,21 @@ expansion_threshold_db = -50.0 # -90..-20 expansion_ratio = 1.5 # 1..4 expansion_range_db = 12.0 # 0..40 -# Static gap trimming: gaps longer than max_gap_s are shortened to gap_pad_s -# (total silence kept, split at the gap edges). Speech is copied through -# unattenuated - no gate, so word attacks are never touched. Use this instead -# of relying on the denoiser to erase room tone in blank sections. -trim_silence = true +# Silence pauses without shortening them. Silero VAD analyzes PRE-denoise +# speech; protected word margins and smooth fades reduce clipped consonants. +# First use installs a small CPU detector and downloads hash-checked weights. +# VAD can miss whispers/breaths; audition the result, not just the meter. +pause_silence = true +pause_threshold = 0.35 # 0.1..0.9; LOWER keeps more possible speech +pause_pre_ms = 120.0 # 30..500; keep this much before detected speech +pause_post_ms = 120.0 # 30..1000; preserve word endings and decay +pause_fade_ms = 10.0 # 2..100; fades outside protected speech +pause_min_ms = 80.0 # 20..1000; minimum zero core, in addition to margins/fades + +# Optional pacing EDIT, not noise removal. Long gaps are shortened to +# gap_pad_s. With pause_silence enabled, only near-zero cores are eligible; +# gap_threshold_db is used only when pause_silence is disabled. +trim_silence = false max_gap_s = 1.5 # 0.5..10; gaps longer than this are shortened gap_pad_s = 0.6 # 0.1..3; total silence kept around each trimmed gap (< max_gap_s) gap_threshold_db = -50.0 # -90..-20; blocks below this count as silence -- cgit v1.2.3