Update app.py

app.py

@@ -1,6 +1,103 @@
-import os 
+import gradio as gr
+from functools import partial
+from pathlib import Path
+import librosa
+import numpy as np
+import matplotlib.pyplot as plt
+import soundfile as sf
+import scipy.signal as sig
+import psola
 
+SEMITONES_IN_OCTAVE = 12
 
+def degrees_from(scale: str):
+    degrees = librosa.key_to_degrees(scale)
+    degrees = np.concatenate((degrees, [degrees[0] + SEMITONES_IN_OCTAVE]))
+    return degrees
 
+def closest_pitch(f0):
+    midi_note = np.around(librosa.hz_to_midi(f0))
+    nan_indices = np.isnan(f0)
+    midi_note[nan_indices] = np.nan
+    return librosa.midi_to_hz(midi_note)
 
-os.system("python runfile.py")
+def closest_pitch_from_scale(f0, scale):
+    if np.isnan(f0):
+        return np.nan
+    degrees = degrees_from(scale)
+    midi_note = librosa.hz_to_midi(f0)
+    degree = midi_note % SEMITONES_IN_OCTAVE
+    degree_id = np.argmin(np.abs(degrees - degree))
+    degree_difference = degree - degrees[degree_id]
+    midi_note -= degree_difference
+    return librosa.midi_to_hz(midi_note)
+
+def aclosest_pitch_from_scale(f0, scale):
+    sanitized_pitch = np.zeros_like(f0)
+    for i in np.arange(f0.shape[0]):
+        sanitized_pitch[i] = closest_pitch_from_scale(f0[i], scale)
+    smoothed_sanitized_pitch = sig.medfilt(sanitized_pitch, kernel_size=11)
+    smoothed_sanitized_pitch[np.isnan(smoothed_sanitized_pitch)] = sanitized_pitch[np.isnan(smoothed_sanitized_pitch)]
+    return smoothed_sanitized_pitch
+
+def autotune(audio, sr, correction_function, plot=False):
+    frame_length = 2048
+    hop_length = frame_length // 4
+    fmin = librosa.note_to_hz('C2')
+    fmax = librosa.note_to_hz('C7')
+
+    f0, voiced_flag, voiced_probabilities = librosa.pyin(audio, frame_length=frame_length, hop_length=hop_length, sr=sr, fmin=fmin, fmax=fmax)
+    corrected_f0 = correction_function(f0)
+
+    if plot:
+        stft = librosa.stft(audio, n_fft=frame_length, hop_length=hop_length)
+        time_points = librosa.times_like(stft, sr=sr, hop_length=hop_length)
+        log_stft = librosa.amplitude_to_db(np.abs(stft), ref=np.max)
+        fig, ax = plt.subplots()
+        img = librosa.display.specshow(log_stft, x_axis='time', y_axis='log', ax=ax, sr=sr, hop_length=hop_length, fmin=fmin, fmax=fmax)
+        fig.colorbar(img, ax=ax, format="%+2.f dB")
+        ax.plot(time_points, f0, label='original pitch', color='cyan', linewidth=2)
+        ax.plot(time_points, corrected_f0, label='corrected pitch', color='orange', linewidth=1)
+        ax.legend(loc='upper right')
+        plt.ylabel('Frequency [Hz]')
+        plt.xlabel('Time [M:SS]')
+        plt.savefig('pitch_correction.png', dpi=300, bbox_inches='tight')
+        plt.close()
+
+    return psola.vocode(audio, sample_rate=int(sr), target_pitch=corrected_f0, fmin=fmin, fmax=fmax)
+
+def process_audio(vocals_file, correction_method, scale, plot):
+    y, sr = librosa.load(vocals_file, sr=None, mono=False)
+    if y.ndim > 1:
+        y = y[0, :]
+    
+    correction_function = closest_pitch if correction_method == 'closest' else partial(aclosest_pitch_from_scale, scale=scale)
+    pitch_corrected_y = autotune(y, sr, correction_function, plot)
+
+    output_file = "pitch_corrected_audio.wav"
+    sf.write(output_file, pitch_corrected_y, sr)
+    
+    if plot:
+        return output_file, 'pitch_correction.png'
+    return output_file, None
+
+def main():
+    with gr.Blocks(title="Hex AutoTune") as demo:
+        gr.Markdown("# Hex Auto-Tune Audio with Pitch Correction")
+        with gr.Row():
+            with gr.Column():
+                vocals_file = gr.Audio(source="upload", type="filepath", label="Upload Vocals File")
+                correction_method = gr.Radio(["closest", "scale"], label="Correction Method", value="closest")
+                scale = gr.Textbox(label="Scale (only for 'scale' method)", placeholder="e.g., C:maj")
+                plot = gr.Checkbox(label="Generate Pitch Correction Plot", value=False)
+                submit = gr.Button("Process")
+            with gr.Column():
+                output_audio = gr.Audio(label="Pitch Corrected Audio")
+                output_image = gr.Image(label="Pitch Correction Plot (if selected)")
+        
+        submit.click(fn=process_audio, inputs=[vocals_file, correction_method, scale, plot], outputs=[output_audio, output_image])
+
+    demo.launch()
+
+if __name__ == '__main__':
+    main()