Update diff_singer_infer.py

diff_singer_infer.py

@@ -1,29 +1,25 @@
-import numpy as np
 import torchaudio
+import numpy as np
 import pyworld as pw
-import soundfile as sf
-from scipy.interpolate import interp1d
+import scipy.io.wavfile as wavfile
 
 def extract_pitch(audio, sr):
-    # Convert audio to double precision for pyworld
-    audio = audio.astype(np.float64)
-
-    _f0, t = pw.dio(audio, sr)
-    f0 = pw.stonemask(audio, _f0, t, sr)
+    _f0, t = pw.dio(audio.astype(np.float64), sr)
+    f0 = pw.stonemask(audio.astype(np.float64), _f0, t, sr)
     return f0
 
-def run_diffsinger_inference(input_audio_path):
-    waveform, sr = torchaudio.load(input_audio_path)
-    audio = waveform[0].numpy()  # Mono
-    f0 = extract_pitch(audio, sr)
+def run_diffsinger_inference(input_path):
+    # Load audio
+    waveform, sr = torchaudio.load(input_path)
+    audio = waveform[0].numpy()
 
-    # Simple manipulation to simulate expressiveness
-    f0_smooth = interp1d(np.arange(len(f0)), f0, kind='cubic', fill_value="extrapolate")(np.linspace(0, len(f0)-1, len(f0)))
-    f0_smooth *= 1.15  # Slight pitch boost
+    # Pitch extraction
+    f0 = extract_pitch(audio, sr)
 
-    # Pitch-shifted output
-    output_audio = pw.synthesize(f0_smooth, pw.cheaptrick(audio.astype(np.float64), f0, np.linspace(0, len(audio)/sr, len(f0)), sr), pw.d4c(audio.astype(np.float64), f0, np.linspace(0, len(audio)/sr, len(f0)), sr), sr)
+    # Simulate pitch & vibrato mod (placeholder until DiffSinger model added)
+    new_audio = audio * 0.8  # just reduce volume for test
 
-    output_path = "output_singing.wav"
-    sf.write(output_path, output_audio, sr)
+    # Save as WAV
+    output_path = "/tmp/output_singing.wav"
+    wavfile.write(output_path, sr, (new_audio * 32767).astype(np.int16))
     return output_path