refactor: replace abstract denoising strength slider with granular controls for post-processing and noise gate parameters

app.py

@@ -151,21 +151,18 @@ def post_process_audio(
     if treble_gain != 0:
         waveform = F_audio.equalizer_biquad(waveform, sample_rate=sr, center_freq=6000.0, gain=treble_gain, Q=0.707)
 
-    # 4. Soft Noise Gate with a floor to keep background noise natural and not too dry
+    # 4. Soft Noise Gate to remove background gating/watery artifacts during silence
     if gate_threshold > -60.0:
         threshold = 10 ** (gate_threshold / 20.0)
         envelope = waveform.abs()
-        # Fast 15ms window to open instantly at the start of words (avoids muddy onset delay)
-        win_size = int(sr * 0.015)
+        # Smooth envelope using 50ms average pooling
+        win_size = int(sr * 0.05)
         if win_size % 2 == 0:
             win_size += 1
         env_padded = torch.nn.functional.pad(envelope, (win_size//2, win_size//2), mode='reflect')
         env_smooth = torch.nn.functional.avg_pool1d(env_padded.unsqueeze(0), kernel_size=win_size, stride=1).squeeze(0)
         # Soft sigmoid gating to prevent clicking
-        raw_gain = torch.sigmoid((env_smooth - threshold) / (threshold * 0.25))
-        # Keep a floor of 35% gain (approx -9dB) so background noise remains continuous & controllable
-        gate_floor = 0.35
-        gain = gate_floor + (1.0 - gate_floor) * raw_gain
+        gain = torch.sigmoid((env_smooth - threshold) / (threshold * 0.25))
         waveform = waveform * gain
 
     # 5. Analog Warmth (Soft Saturation / Tube Limiting)
@@ -185,29 +182,25 @@ def demo_fn(
     speech_upl: str,
     noise_type: str,
     snr: int,
-    denoising_strength: float,
-    vocal_polish: bool,
+    atten_lim_db: float,
+    wet_dry_mix: float,
+    post_process: bool,
+    low_cut_freq: float,
+    bass_gain: float,
+    treble_gain: float,
+    gate_threshold: float,
+    enable_warmth: bool,
     mic_input: Optional[str] = None,
 ):
     if input_type == "mic":
         speech_upl = mic_input
 
     sr = config("sr", 48000, int, section="df")
-    logger.info(f"Got parameters input_type: {input_type}, speech_upl: {speech_upl}, noise: {noise_type}, snr: {snr}, strength: {denoising_strength}, polish: {vocal_polish}")
+    logger.info(f"Got parameters input_type: {input_type}, speech_upl: {speech_upl}, noise: {noise_type}, snr: {snr}, atten_lim_db: {atten_lim_db}, wet_dry_mix: {wet_dry_mix}, post_process: {post_process}, low_cut: {low_cut_freq}, bass: {bass_gain}, treble: {treble_gain}, gate: {gate_threshold}, warmth: {enable_warmth}")
     snr = int(snr)
     noise_fn = NOISES[noise_type]
     meta = AudioMetaData(-1, -1, -1, -1, "")
 
-    # Map the simple 0-100% denoising strength to advanced model parameters
-    if denoising_strength == 0:
-        atten_lim_db = 0.0
-        wet_dry_mix = 0.0
-    else:
-        # Interpolate maximum attenuation between 6dB and 30dB
-        atten_lim_db = 6.0 + (denoising_strength / 100.0) * 24.0
-        # Interpolate dry/wet mix between 50% and 100%
-        wet_dry_mix = 50.0 + (denoising_strength / 100.0) * 50.0
-
     max_s = 3600  # allow up to 1 hour (3600 seconds)
 
     if speech_upl is not None:
@@ -249,18 +242,9 @@ def demo_fn(
     alpha = wet_dry_mix / 100.0
     enhanced = alpha * enhanced + (1 - alpha) * sample
 
-    # Apply professional post-processing (Low-cut filter + bass EQ + treble EQ + gate + warmth) using optimized studio defaults
-    if vocal_polish:
-        # We set low_cut_freq=110.0 to cut muddy frequency spikes at the start of words
-        enhanced = post_process_audio(
-            enhanced,
-            sr,
-            low_cut_freq=110.0,
-            bass_gain=0.0,
-            treble_gain=3.0,
-            gate_threshold=-45.0,
-            enable_warmth=True
-        )
+    # Apply professional post-processing (Low-cut filter + bass EQ + treble EQ + gate + warmth)
+    if post_process:
+        enhanced = post_process_audio(enhanced, sr, low_cut_freq, bass_gain, treble_gain, gate_threshold, enable_warmth)
 
     lim = torch.linspace(0.0, 1.0, int(sr * 0.15)).unsqueeze(0)
     lim = torch.cat((lim, torch.ones(1, enhanced.shape[-1] - lim.shape[1])), dim=1)
@@ -441,14 +425,53 @@ with gr.Blocks() as demo:
                     value="10",
                 ),
                 gr.Slider(
-                    minimum=0,
+                    minimum=6,
+                    maximum=35,
+                    step=1,
+                    value=15,
+                    label="Max Attenuation (dB) - lower is more natural/clear, higher reduces more noise",
+                ),
+                gr.Slider(
+                    minimum=50,
                     maximum=100,
                     step=5,
+                    value=90,
+                    label="Voice Naturalness Mix (%) - 100% is fully denoised, 90% blends back some original voice",
+                ),
+                gr.Checkbox(
+                    label="Enable Post-Processing EQ & Filters",
+                    value=True,
+                ),
+                gr.Slider(
+                    minimum=20,
+                    maximum=200,
+                    step=10,
                     value=80,
-                    label="Denoising Strength (%) - higher removes more noise, lower is more natural",
+                    label="Low-Cut Filter Cutoff (Hz) - cuts muddy AC rumble/handling noise",
+                ),
+                gr.Slider(
+                    minimum=-10,
+                    maximum=10,
+                    step=1,
+                    value=0,
+                    label="Voice Bass EQ Boost (dB) at 150Hz - adds warmth",
+                ),
+                gr.Slider(
+                    minimum=-10,
+                    maximum=10,
+                    step=1,
+                    value=3,
+                    label="Voice Treble EQ Boost (dB) at 6kHz - adds crispness/air",
+                ),
+                gr.Slider(
+                    minimum=-60,
+                    maximum=-30,
+                    step=5,
+                    value=-45,
+                    label="Noise Gate Threshold (dB) - lower keeps quiet sound, higher silences gaps",
                 ),
                 gr.Checkbox(
-                    label="Vocal Polish (80Hz Low-Cut, Treble EQ & Analog Warmth)",
+                    label="Enable Analog Warmth (Soft Saturation)",
                     value=True,
                 ),
                 mic_input,
@@ -467,10 +490,10 @@ with gr.Blocks() as demo:
     radio.change(toggle, radio, [mic_input, audio_file])
     gr.Examples(
         [
-            ["file", "./samples/p232_013_clean.wav", "Kitchen", "10", 80, True, None],
-            ["file", "./samples/p232_013_clean.wav", "Cafe", "10", 80, True, None],
-            ["file", "./samples/p232_019_clean.wav", "Cafe", "10", 80, True, None],
-            ["file", "./samples/p232_019_clean.wav", "River", "10", 80, True, None],
+            ["file", "./samples/p232_013_clean.wav", "Kitchen", "10", 15, 90, True, 80, 0, 3, -45, True, None],
+            ["file", "./samples/p232_013_clean.wav", "Cafe", "10", 15, 90, True, 80, 0, 3, -45, True, None],
+            ["file", "./samples/p232_019_clean.wav", "Cafe", "10", 15, 90, True, 80, 0, 3, -45, True, None],
+            ["file", "./samples/p232_019_clean.wav", "River", "10", 15, 90, True, 80, 0, 3, -45, True, None],
         ],
         fn=demo_fn,
         inputs=inputs,