feat: add optional DSP post-processing for audio clarification and normalization

app.py

@@ -120,19 +120,38 @@ def ensure_wav(filepath: str) -> str:
     return filepath
 
 
+import torchaudio.functional as F_audio
+
+def post_process_audio(waveform: Tensor, sr: int) -> Tensor:
+    """Apply professional DSP filters to make voice audio crisp and clear."""
+    # 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)
+
+    return waveform
+
+
 def demo_fn(
     speech_upl: str,
     noise_type: str,
     snr: int,
     atten_lim_db: float,
     wet_dry_mix: float,
+    post_process: bool,
     mic_input: Optional[str] = None,
 ):
     if mic_input:
         speech_upl = mic_input
 
     sr = config("sr", 48000, int, section="df")
-    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}")
+    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, "")
@@ -177,8 +196,12 @@ def demo_fn(
     alpha = wet_dry_mix / 100.0
     enhanced = alpha * enhanced + (1 - alpha) * sample
 
+    # Apply professional post-processing (Low-cut filter + presence boost EQ)
+    if post_process:
+        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)
+    lim = torch.cat((lim, torch.ones(1, enhanced.shape[-1] - lim.shape[1])), dim=1)
     enhanced = enhanced * lim
 
     if meta.sample_rate != sr:
@@ -368,6 +391,10 @@ with gr.Blocks() as demo:
                     value=90,
                     label="Voice Naturalness Mix (%) - 100% is fully denoised, 90% blends back some original voice",
                 ),
+                gr.Checkbox(
+                    label="Post-Process (80Hz Low-Cut & Presence Boost)",
+                    value=True,
+                ),
                 mic_input,
             ]
             btn = gr.Button("Generate")
@@ -384,10 +411,10 @@ with gr.Blocks() as demo:
     radio.change(toggle, radio, [mic_input, audio_file])
     gr.Examples(
         [
-            ["./samples/p232_013_clean.wav", "Kitchen", "10", 15, 90, None],
-            ["./samples/p232_013_clean.wav", "Cafe", "10", 15, 90, None],
-            ["./samples/p232_019_clean.wav", "Cafe", "10", 15, 90, None],
-            ["./samples/p232_019_clean.wav", "River", "10", 15, 90, 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,