Fix audio distortion by normalizing to -1 dBFS with headroom

Peak-normalize diffusion output instead of hard-clamping to [-1,1],
which was chopping peaks and causing audible clipping distortion.
Also save as float32 WAV to let torchaudio handle encoding safely.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

app.py

@@ -132,18 +132,20 @@ def generate_audio(prompt, negative_prompt, bars, bpm, note, scale, steps, cfg_s
         
         # 1. Rearrange to [channels, samples] and ensure float32
         audio = rearrange(audio, "b d n -> d (b n)").to(torch.float32)
-        
-        # 2. PEAK NORMALIZATION (The RC Repo Fix)
-        # This scales the entire clip so the highest peak is exactly 1.0/-1.0.
-        # This prevents the 'overdriven' sound by keeping all data within the valid range.
+
+        # 2. PEAK NORMALIZATION — prevents clipping distortion
+        # The diffusion model outputs can far exceed [-1, 1].
+        # Hard-clamping (the repo default) chops those peaks → audible distortion.
+        # Instead, scale the whole waveform so the loudest peak sits at -1 dBFS
+        # (≈0.89), leaving headroom to avoid inter-sample clipping in DAWs.
         max_amp = torch.abs(audio).max()
-        if max_amp > 0:
-            audio = audio / max_amp
-            
+        if max_amp > 1e-8:
+            audio = audio / max_amp * 0.89125  # -1 dBFS headroom
+
         # 3. Trim to the exact deterministic grid length
         end = min(int(audio.shape[-1]), int(clip_samples))
         audio = audio[:, :max(1, end)].contiguous()
-        
+
         # 4. Apply a tiny 15ms fade-out to avoid clicks at the end
         fade_ms = 15.0
         fade_len = int(round((fade_ms / 1000.0) * SAMPLE_RATE))
@@ -151,15 +153,14 @@ def generate_audio(prompt, negative_prompt, bars, bpm, note, scale, steps, cfg_s
             fade_len = min(fade_len, audio.shape[-1])
             ramp = torch.linspace(1.0, 0.0, steps=fade_len, device=audio.device)
             audio[:, -fade_len:] *= ramp
-            
-        # 5. Convert to 16-bit PCM WAV (Standard for Stable Audio outputs)
-        # We clamp just to be safe, though normalization should already have handled it.
-        wav_i16 = (audio.clamp(-1, 1) * 32767.0).to(torch.int16).cpu()
-        
+
+        # 5. Save as float32 WAV — let torchaudio handle bit-depth encoding
+        audio = audio.clamp(-1, 1).cpu()
+
         with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
             output_path = tmp.name
-            
-        torchaudio.save(output_path, wav_i16, SAMPLE_RATE)
+
+        torchaudio.save(output_path, audio, SAMPLE_RATE)
         return output_path
         
     except Exception as e: