From 1bbe59f78be9a8355bb0bee2bc8ab14ca42ac71c Mon Sep 17 00:00:00 2001 From: historia Date: Fri, 11 Sep 2026 22:03:43 -0400 Subject: Add ClearVoice MossFormer2 denoiser backend and static gap trimming --- config.toml | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) (limited to 'config.toml') diff --git a/config.toml b/config.toml index a574717..75e0e3b 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 = "deepfilter" # DeepFilterNet3 only; fft is explicit non-AI fallback; none +denoiser = "clearvoice" # clearvoice (MossFormer2_SE_48K, best attack preservation), deepfilter; fft is explicit non-AI fallback; none device = "auto" # auto, cpu, cuda; explicit cuda fails if unavailable -denoise_strength = 0.85 # 0..1; AI dry/wet blend; zero skips denoising and AI setup +denoise_strength = 1.0 # 0..1; AI dry/wet blend; zero skips denoising and AI setup fft_reduction_db = 8.0 # 0.01..30; FFT reduction is scaled by denoise_strength highpass = true @@ -26,6 +26,15 @@ 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 +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 + eq = true warmth_db = 0.5 # -12..12; 140 Hz mud_db = -1.5 # -12..12; 300 Hz -- cgit v1.2.3