feat: add soft noise gate and analog warmth DSP processing to post-processing pipeline

app.py

@@ -135,6 +135,8 @@ def post_process_audio(
     low_cut_freq: float,
     bass_gain: float,
     treble_gain: float,
+    gate_threshold: float,
+    enable_warmth: bool,
 ) -> Tensor:
     """Apply professional DSP filters to make voice audio crisp and clear."""
     # 1. High-pass filter (Low-cut) to cut sub-bass rumble and AC hums
@@ -149,7 +151,25 @@ 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. Normalize peak to -1.0 dBFS (0.9 max amplitude) to maximize volume without clipping
+    # 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()
+        # 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
+        gain = torch.sigmoid((env_smooth - threshold) / (threshold * 0.25))
+        waveform = waveform * gain
+
+    # 5. Analog Warmth (Soft Saturation / Tube Limiting)
+    if enable_warmth:
+        waveform = torch.tanh(waveform * 1.05) / 1.05
+
+    # 6. Normalize peak to -1.0 dBFS (0.9 max amplitude) to maximize volume without clipping
     max_val = waveform.abs().max()
     if max_val > 0:
         waveform = waveform * (0.9 / max_val)
@@ -168,13 +188,15 @@ def demo_fn(
     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}, 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}")
+    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, "")
@@ -220,9 +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)
+    # 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)
+        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)
@@ -441,6 +463,17 @@ with gr.Blocks() as demo:
                     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="Enable Analog Warmth (Soft Saturation)",
+                    value=True,
+                ),
                 mic_input,
             ]
             btn = gr.Button("Generate")
@@ -457,10 +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", 15, 90, True, 80, 0, 3, None],
-            ["file", "./samples/p232_013_clean.wav", "Cafe", "10", 15, 90, True, 80, 0, 3, None],
-            ["file", "./samples/p232_019_clean.wav", "Cafe", "10", 15, 90, True, 80, 0, 3, None],
-            ["file", "./samples/p232_019_clean.wav", "River", "10", 15, 90, True, 80, 0, 3, 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,