feat: chunking, zipenhancer denoising

1 files changed, 18 insertions, 4 deletions

diff --git a/config.toml b/config.toml
index 412ba17..1f05b75 100644
--- a/config.toml
+++ b/config.toml
@@ -9,8 +9,8 @@ profile = "radio"
 # Output container/codec: wav | flac | mp3          (default: wav)
 # format = "wav"
 
-# Bit depth: 16 | 24 | 32                           (default: 16)
-# bit_depth = 16
+# Bit depth: 16 | 24 | 32                           (default: 32)
+# bit_depth = 32
 
 # Output sample rate in Hz                          (default: 44100 audiobook / 48000 podcast)
 # sample_rate = 44100
@@ -18,6 +18,13 @@ profile = "radio"
 # Compute device: auto (CUDA if present) | cuda | cpu
 # device = "auto"
 
+# GPU engine processing: engines receive the whole file by default so models
+# run one warm pass with no seam artifacts and a single model cold-start.
+# chunk > 0 forces chunked streaming in seconds (very long files / low VRAM);
+# whole-file runs fall back to chunks automatically if they run out of memory.
+# engine_chunk = 0.0     # 0 = whole file (default)
+# engine_overlap = 0.5   # crossfade between chunks in seconds
+
 # Loudness target and true-peak ceiling in dB.
 # Without these, each profile uses RMS -20 / ceiling -3 (audiobook) or -16 LUFS / -1.5 (podcast).
 # target = -20.0
@@ -29,11 +36,17 @@ profile = "radio"
 
 # Denoiser: "dfn3" (DeepFilterNet3, default, fast, 48 kHz full-band)
 #           "zipenhancer" (ICASSP 2025 SOTA, 16 kHz native, bandwidth restored)
+#           "spectral" (pure-DSP dd-wiener with a file-global noise profile,
+#                       no downloads, no wobble)
 #           "off" (pure DSP, no heavy downloads)
 # Heavy deps + model weights are installed lazily on first use.
 [denoise]
 engine = "dfn3"
-strength = 1.0
+# 0.9 flattens residual denoiser gain wobble (volume pumping); the small dry
+# blend trades a whisper of noise back for level stability. 1.0 = max attack.
+strength = 0.9
+# pf = false  # DeepFilterNet post filter: extra noise reduction, opt-in —
+#             # may eat soft speech on clean recordings (off by default)
 
 # Speech enhancement: "off" (default), "mossformer2" (studio restoration),
 #                     "resemble" (generative, opt-in isolated venv, may alter timbre)
@@ -52,7 +65,8 @@ engine = "off"
 # deess    = de-esser active around 5.5-8 kHz
 # presence = boost around 3 kHz
 # air      = high shelf from 9-10 kHz
-# breath   = expander depth used to duck breaths
+# breath   = downward expander depth used to duck breaths and residual hiss
+#            in sentence gaps (0 disables; up to ~24 dB of ducking at 1.0)
 
 [audiobook]
 mud = 0.8