refactor: simplify audio post-processing to fixed filters and remove redundant UI controls

app.py

@@ -129,47 +129,15 @@ def ensure_wav(filepath: str) -> str:
 
 import torchaudio.functional as F_audio
 
-def post_process_audio(
-    waveform: Tensor,
-    sr: int,
-    low_cut_freq: float,
-    bass_gain: float,
-    treble_gain: float,
-    gate_threshold: float,
-    enable_warmth: bool,
-) -> Tensor:
+def post_process_audio(waveform: Tensor, sr: int) -> 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
-    if low_cut_freq > 20:
-        waveform = F_audio.highpass_biquad(waveform, sample_rate=sr, cutoff_freq=low_cut_freq)
-
-    # 2. Bass EQ peaking filter at 150 Hz to add warmth or reduce muddiness
-    if bass_gain != 0:
-        waveform = F_audio.equalizer_biquad(waveform, sample_rate=sr, center_freq=150.0, gain=bass_gain, Q=0.707)
-
-    # 3. Treble EQ peaking filter at 6000 Hz to add air, presence, and vocal crispness
-    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 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
+    # 1. High-pass filter at 80 Hz to cut muddy sub-bass rumble and AC hums
+    waveform = F_audio.highpass_biquad(waveform, sample_rate=sr, cutoff_freq=80.0)
+
+    # 2. Treble peaking equalizer at 6000 Hz (+3.0 dB) to add air, presence, and vocal crispness
+    waveform = F_audio.equalizer_biquad(waveform, sample_rate=sr, center_freq=6000.0, gain=3.0, Q=0.707)
+
+    # 3. 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)
@@ -178,25 +146,19 @@ def post_process_audio(
 
 
 def demo_fn(
-    input_type: str,
     speech_upl: str,
     noise_type: str,
     snr: int,
     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":
+    if mic_input:
         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}, gate: {gate_threshold}, warmth: {enable_warmth}")
+    logger.info(f"Got parameters 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}")
     snr = int(snr)
     noise_fn = NOISES[noise_type]
     meta = AudioMetaData(-1, -1, -1, -1, "")
@@ -242,9 +204,9 @@ 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)
+    # Apply professional post-processing (Low-cut filter + presence boost EQ)
     if post_process:
-        enhanced = post_process_audio(enhanced, sr, low_cut_freq, bass_gain, treble_gain, gate_threshold, enable_warmth)
+        enhanced = post_process_audio(enhanced, sr)
 
     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)
@@ -412,7 +374,6 @@ with gr.Blocks() as demo:
             mic_input = gr.Mic(label="Input", type="filepath", visible=False)
             audio_file = gr.Audio(type="filepath", label="Input", visible=True)
             inputs = [
-                radio,
                 audio_file,
                 gr.Dropdown(
                     label="Add background noise",
@@ -439,39 +400,7 @@ with gr.Blocks() as demo:
                     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="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="Enable Analog Warmth (Soft Saturation)",
+                    label="Post-Process (80Hz Low-Cut & Presence Boost)",
                     value=True,
                 ),
                 mic_input,
@@ -490,10 +419,10 @@ 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, -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],
+            ["./samples/p232_013_clean.wav", "Kitchen", "10", 15, 90, True, None],
+            ["./samples/p232_013_clean.wav", "Cafe", "10", 15, 90, True, None],
+            ["./samples/p232_019_clean.wav", "Cafe", "10", 15, 90, True, None],
+            ["./samples/p232_019_clean.wav", "River", "10", 15, 90, True, None],
         ],
         fn=demo_fn,
         inputs=inputs,