Create app.py

app.py

@@ -0,0 +1,385 @@
+import gradio as gr
+import os
+import shutil
+import asyncio
+import librosa
+import librosa.display
+import soundfile as sf
+import numpy as np
+import time
+import zipfile
+import tempfile
+import matplotlib.pyplot as plt
+import matplotlib
+matplotlib.use('Agg')
+
+temp_files = []
+
+def cleanup_temp_files():
+    global temp_files
+    for file_path in temp_files:
+        if os.path.exists(file_path):
+            os.remove(file_path)
+    temp_files = []
+
+def update_output_visibility(choice):
+    if "2 Stems" in choice:
+        return {
+            vocals_output: gr.update(visible=True),
+            drums_output: gr.update(visible=False),
+            bass_output: gr.update(visible=False),
+            other_output: gr.update(visible=True, label="Instrumental (No Vocals)")
+        }
+    elif "4 Stems" in choice:
+        return {
+            vocals_output: gr.update(visible=True),
+            drums_output: gr.update(visible=True),
+            bass_output: gr.update(visible=True),
+            other_output: gr.update(visible=True, label="Other")
+        }
+
+async def separate_stems(audio_file_path, stem_choice, progress=gr.Progress(track_tqdm=True)):
+    if audio_file_path is None: raise gr.Error("No audio file uploaded!")
+    progress(0, desc="Starting...")
+    try:
+        progress(0.05, desc="Preparing audio file...")
+        original_filename_base = os.path.basename(audio_file_path).rsplit('.', 1)[0]
+        stable_input_path = f"stable_input_{original_filename_base}.wav"
+        shutil.copy(audio_file_path, stable_input_path)
+
+        model_arg = "--two-stems=vocals" if "2 Stems" in stem_choice else ""
+        output_dir = "separated"
+        if os.path.exists(output_dir): shutil.rmtree(output_dir)
+
+        command = f"python3 -m demucs {model_arg} -o \"{output_dir}\" \"{stable_input_path}\""
+        progress(0.2, desc="Running Demucs (this can take a minute)...")
+
+        process = await asyncio.create_subprocess_shell(
+            command,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE)
+
+        stdout, stderr = await process.communicate()
+
+        if process.returncode != 0:
+            raise gr.Error(f"Demucs failed to run. Error: {stderr.decode()[:500]}")
+
+        progress(0.8, desc="Locating separated stem files...")
+        stable_filename_base = os.path.basename(stable_input_path).rsplit('.', 1)[0]
+        model_folder_name = next(os.walk(output_dir))[1][0]
+        stems_path = os.path.join(output_dir, model_folder_name, stable_filename_base)
+
+        if not os.path.exists(stems_path):
+            raise gr.Error(f"Demucs finished, but the output directory was not found!")
+
+        vocals_path = os.path.join(stems_path, "vocals.wav") if os.path.exists(os.path.join(stems_path, "vocals.wav")) else None
+        drums_path = os.path.join(stems_path, "drums.wav") if os.path.exists(os.path.join(stems_path, "drums.wav")) else None
+        bass_path = os.path.join(stems_path, "bass.wav") if os.path.exists(os.path.join(stems_path, "bass.wav")) else None
+        other_filename = "no_vocals.wav" if "2 Stems" in stem_choice else "other.wav"
+        other_path = os.path.join(stems_path, other_filename) if os.path.exists(os.path.join(stems_path, other_filename)) else None
+
+        os.remove(stable_input_path)
+
+        # Detect bars for each stem after separation
+        vocals_bar_times = None
+        drums_bar_times = None
+        bass_bar_times = None
+        other_bar_times = None
+
+        if vocals_path:
+            vocals_audio_data = sf.read(vocals_path)
+            _, _, vocals_bar_times = detect_bars(vocals_audio_data)
+        if drums_path:
+            drums_audio_data = sf.read(drums_path)
+            _, _, drums_bar_times = detect_bars(drums_audio_data)
+        if bass_path:
+            bass_audio_data = sf.read(bass_path)
+            _, _, bass_bar_times = detect_bars(bass_audio_data)
+        if other_path:
+            other_audio_data = sf.read(other_path)
+            _, _, other_bar_times = detect_bars(other_audio_data)
+
+
+        return vocals_path, drums_path, bass_path, other_path, vocals_bar_times, drums_bar_times, bass_bar_times, other_bar_times
+
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        raise gr.Error(str(e))
+
+def visualize_slices(stem_audio_data, sensitivity, progress=gr.Progress(track_tqdm=True)):
+    if stem_audio_data is None:
+        gr.Warning("This stem is empty. Cannot visualize.")
+        return None, None, None
+
+    sample_rate, y_int = stem_audio_data
+    y = librosa.util.buf_to_float(y_int)
+
+    progress(0.3, desc="Finding transients...")
+    onset_frames = librosa.onset.onset_detect(y=librosa.to_mono(y.T) if y.ndim > 1 else y, sr=sample_rate, wait=1, pre_avg=1, post_avg=1, post_max=1, delta=sensitivity)
+    onset_times = librosa.frames_to_time(onset_frames, sr=sample_rate)
+
+    progress(0.7, desc="Generating waveform plot...")
+    fig, ax = plt.subplots(figsize=(10, 3))
+    fig.patch.set_facecolor('#1f2937')
+    ax.set_facecolor('#111827')
+    librosa.display.waveshow(y, sr=sample_rate, ax=ax, color='#32f6ff', alpha=0.7)
+    for t in onset_times:
+        ax.axvline(x=t, color='#ff3b3b', linestyle='--', linewidth=1)
+    ax.tick_params(colors='gray'); ax.xaxis.label.set_color('gray'); ax.yaxis.label.set_color('gray')
+    ax.set_xlabel("Time (s)"); ax.set_ylabel("Amplitude"); ax.set_title("Detected Slices", color='white')
+    plt.tight_layout()
+
+    progress(1, desc="Done!")
+    return fig, onset_times, stem_audio_data
+
+def preview_slice(active_stem_audio, onset_times, evt: gr.SelectData):
+    if active_stem_audio is None or onset_times is None: return None
+    sample_rate, y = active_stem_audio
+
+    # Convert click event coordinates to time
+    # evt.index[0] is the x-coordinate of the click in pixels
+    # evt.target[0] is the width of the plot in pixels
+    # len(y) / sample_rate is the total duration of the audio in seconds
+    clicked_time = evt.index[0] * (len(y) / sample_rate) / evt.target[0]
+
+    start_time = 0
+    end_time = len(y) / sample_rate
+
+    # Find the closest onset time before the clicked time
+    onsets_before = onset_times[onset_times <= clicked_time]
+    if len(onsets_before) > 0:
+        start_time = onsets_before[-1]
+
+    # Find the closest onset time after the clicked time
+    onsets_after = onset_times[onset_times > clicked_time]
+    if len(onsets_after) > 0:
+        end_time = onsets_after[0]
+    else:
+        # If no onset after the clicked time, slice to the end of the audio
+        end_time = len(y) / sample_rate
+
+
+    start_sample = librosa.time_to_samples(start_time, sr=sample_rate)
+    end_sample = librosa.time_to_samples(end_time, sr=sample_rate)
+
+    # Ensure start_sample is less than end_sample
+    if start_sample >= end_sample:
+        # If the click is exactly on or after the last onset, preview a small segment at the end
+        if len(onset_times) > 0:
+             start_sample = librosa.time_to_samples(onset_times[-1], sr=sample_rate)
+             end_sample = len(y)
+        else:
+             # If no onsets detected, slice the whole audio
+             start_sample = 0
+             end_sample = len(y)
+
+
+    sliced_audio = y[start_sample:end_sample]
+    return (sample_rate, sliced_audio)
+
+def download_slice(sliced_audio_data):
+    if sliced_audio_data is None:
+        gr.Warning("No slice preview available to download.")
+        return None
+
+    sample_rate, y = sliced_audio_data
+
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False, prefix="slice_") as tmp_file:
+        sf.write(tmp_file.name, y, sample_rate)
+        global temp_files
+        temp_files.append(tmp_file.name)
+        return tmp_file.name
+
+def detect_bars(stem_audio_data):
+    if stem_audio_data is None:
+        return None, None, None
+
+    sample_rate, y_int = stem_audio_data
+    y = librosa.util.buf_to_float(y_int)
+    y_mono = librosa.to_mono(y.T) if y.ndim > 1 else y
+
+    # Estimate tempo
+    tempo, beat_frames = librosa.beat.beat_track(y=y_mono, sr=sample_rate)
+
+    # Convert beat frames to beat times
+    beat_times = librosa.frames_to_time(beat_frames, sr=sample_rate)
+
+    # Calculate bar times (assuming 4 beats per bar)
+    bar_times = beat_times[::4]
+
+    return tempo, beat_times, bar_times
+
+def create_loop(stem_audio_data, bar_times, loop_length):
+    if stem_audio_data is None or bar_times is None or len(bar_times) < 2:
+        gr.Warning("Insufficient data to create a loop.")
+        return None
+
+    sample_rate, y_int = stem_audio_data
+    y = librosa.util.buf_to_float(y_int)
+    y_mono = librosa.to_mono(y.T) if y.ndim > 1 else y
+
+    # Parse loop length
+    num_bars = int(loop_length.split(" ")[0])
+
+    # Find the start of the first full bar (assuming bar_times[0] is the start of the first bar)
+    # If we want to start from the beginning of the audio, we can use 0 as the start time.
+    # For now, let's assume we start from the first detected bar.
+    start_time = bar_times[0]
+
+    # Calculate the duration of one bar
+    bar_duration = bar_times[1] - bar_times[0] if len(bar_times) > 1 else 0
+
+    # Calculate the end time for the loop
+    end_time = start_time + (num_bars * bar_duration)
+
+    # Ensure the end time does not exceed the audio duration
+    audio_duration = len(y) / sample_rate
+    end_time = min(end_time, audio_duration)
+
+    # Convert times to samples
+    start_sample = librosa.time_to_samples(start_time, sr=sample_rate)
+    end_sample = librosa.time_to_samples(end_time, sr=sample_rate)
+
+    # Extract the loop segment
+    looped_audio = y_mono[start_sample:end_sample]
+
+    # Save the looped audio to a temporary file
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False, prefix="loop_") as tmp_file:
+        sf.write(tmp_file.name, looped_audio, sample_rate)
+        global temp_files
+        temp_files.append(tmp_file.name)
+        return tmp_file.name
+
+def cut_all_oneshots(stem_audio_data, onset_times):
+    if stem_audio_data is None or onset_times is None or len(onset_times) < 1:
+        gr.Warning("Insufficient data or onsets detected to cut one-shots.")
+        return None
+
+    sample_rate, y_int = stem_audio_data
+    y = librosa.util.buf_to_float(y_int)
+    y_mono = librosa.to_mono(y.T) if y.ndim > 1 else y
+
+    oneshot_files = []
+    audio_duration = len(y_mono) / sample_rate
+
+    for i in range(len(onset_times)):
+        start_time = onset_times[i]
+        end_time = onset_times[i+1] if i < len(onset_times) - 1 else audio_duration
+
+        start_sample = librosa.time_to_samples(start_time, sr=sample_rate)
+        end_sample = librosa.time_to_samples(end_time, sr=sample_rate)
+
+        # Ensure start_sample is less than end_sample, add a small buffer if necessary
+        if start_sample >= end_sample:
+            end_sample = start_sample + int(0.01 * sample_rate) # Add 10ms buffer if start is equal to or after end
+            if end_sample > len(y_mono):
+                 end_sample = len(y_mono)
+
+        segment = y_mono[start_sample:end_sample]
+
+        # Save each segment to a temporary file
+        with tempfile.NamedTemporaryFile(suffix=f"_{i}.wav", delete=False, prefix="oneshot_") as tmp_file:
+            sf.write(tmp_file.name, segment, sample_rate)
+            oneshot_files.append(tmp_file.name)
+
+    if not oneshot_files:
+        gr.Warning("No one-shots were successfully cut.")
+        return None
+
+    # Create a zip archive of the temporary one-shot files
+    with tempfile.NamedTemporaryFile(suffix=".zip", delete=False, prefix="oneshots_archive_") as zip_file:
+        with zipfile.ZipFile(zip_file.name, 'w') as zipf:
+            for file_path in oneshot_files:
+                zipf.write(file_path, os.path.basename(file_path))
+
+    # Add the zip file and individual oneshot files to the temp_files list for cleanup
+    global temp_files
+    temp_files.extend(oneshot_files)
+    temp_files.append(zip_file.name)
+
+    return zip_file.name
+
+
+with gr.Blocks(theme=gr.themes.Default(primary_hue="blue", secondary_hue="red")) as demo:
+    gr.Markdown("# 🎵 Loop Architect")
+    onset_times_state = gr.State(value=None)
+    active_stem_state = gr.State(value=None)
+    vocals_bar_times_state = gr.State(value=None)
+    drums_bar_times_state = gr.State(value=None)
+    bass_bar_times_state = gr.State(value=None)
+    other_bar_times_state = gr.State(value=None)
+
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            gr.Markdown("### 1. Separate Stems")
+            audio_input = gr.Audio(type="filepath", label="Upload a Track")
+            stem_options = gr.Radio(["4 Stems (Vocals, Drums, Bass, Other)", "2 Stems (Vocals + Instrumental)"], label="Separation Type", value="4 Stems (Vocals, Drums, Bass, Other)")
+            submit_button = gr.Button("Separate Stems")
+
+        with gr.Column(scale=2):
+            with gr.Accordion("Separated Stems", open=True):
+                with gr.Row():
+                    vocals_output = gr.Audio(label="Vocals", scale=2)
+                    with gr.Column(scale=1):
+                        slice_vocals_btn = gr.Button("Visualize Slices")
+                        vocals_loop_length = gr.Dropdown(choices=["4 Bars", "8 Bars", "16 Bars"], label="Loop Length", value="4 Bars")
+                        create_vocals_loop_btn = gr.Button("Create Loop")
+                    vocals_loop_output = gr.Audio(label="Vocals Loop", visible=False, scale=2)
+                    vocals_loop_download_btn = gr.DownloadButton(value="Download Loop", visible=False)
+                with gr.Row():
+                    drums_output = gr.Audio(label="Drums", scale=2)
+                    with gr.Column(scale=1):
+                        slice_drums_btn = gr.Button("Visualize Slices")
+                        drums_loop_length = gr.Dropdown(choices=["4 Bars", "8 Bars", "16 Bars"], label="Loop Length", value="4 Bars")
+                        create_drums_loop_btn = gr.Button("Create Loop")
+                    drums_loop_output = gr.Audio(label="Drums Loop", visible=False, scale=2)
+                    drums_loop_download_btn = gr.DownloadButton(value="Download Loop", visible=False)
+                with gr.Row():
+                    bass_output = gr.Audio(label="Bass", scale=2)
+                    with gr.Column(scale=1):
+                        slice_bass_btn = gr.Button("Visualize Slices")
+                        bass_loop_length = gr.Dropdown(choices=["4 Bars", "8 Bars", "16 Bars"], label="Loop Length", value="4 Bars")
+                        create_bass_loop_btn = gr.Button("Create Loop")
+                    bass_loop_output = gr.Audio(label="Bass Loop", visible=False, scale=2)
+                    bass_loop_download_btn = gr.DownloadButton(value="Download Loop", visible=False)
+                with gr.Row():
+                    other_output = gr.Audio(label="Other / Instrumental", scale=2)
+                    with gr.Column(scale=1):
+                        slice_other_btn = gr.Button("Visualize Slices")
+                        other_loop_length = gr.Dropdown(choices=["4 Bars", "8 Bars", "16 Bars"], label="Loop Length", value="4 Bars")
+                        create_other_loop_btn = gr.Button("Create Loop")
+                    other_loop_output = gr.Audio(label="Other Loop", visible=False, scale=2)
+                    other_loop_download_btn = gr.DownloadButton(value="Download Loop", visible=False)
+
+
+            gr.Markdown("### Slice Editor")
+            sensitivity_slider = gr.Slider(minimum=0, maximum=1, value=0.5, label="Onset Sensitivity")
+            slice_plot = gr.Image(label="Click a region on the waveform to preview a slice")
+            preview_player = gr.Audio(label="Slice Preview")
+            download_slice_btn = gr.DownloadButton(value="Download Slice", visible=False)
+            cut_all_oneshots_btn = gr.Button(value="Cut All Oneshots")
+            cut_oneshots_download_btn = gr.DownloadButton(value="Download All Oneshots", visible=False)
+
+
+    audio_input.change(fn=cleanup_temp_files)
+    submit_button.click(fn=separate_stems, inputs=[audio_input, stem_options], outputs=[vocals_output, drums_output, bass_output, other_output, vocals_bar_times_state, drums_bar_times_state, bass_bar_times_state, other_bar_times_state])
+    stem_options.change(fn=update_output_visibility, inputs=stem_options, outputs=[vocals_output, drums_output, bass_output, other_output])
+
+    slice_vocals_btn.click(fn=visualize_slices, inputs=[vocals_output, sensitivity_slider], outputs=[slice_plot, onset_times_state, active_stem_state])
+    slice_drums_btn.click(fn=visualize_slices, inputs=[drums_output, sensitivity_slider], outputs=[slice_plot, onset_times_state, active_stem_state])
+    slice_bass_btn.click(fn=visualize_slices, inputs=[bass_output, sensitivity_slider], outputs=[slice_plot, onset_times_state, active_stem_state])
+    slice_other_btn.click(fn=visualize_slices, inputs=[other_output, sensitivity_slider], outputs=[slice_plot, onset_times_state, active_stem_state])
+
+    slice_plot.select(fn=preview_slice, inputs=[active_stem_state, onset_times_state], outputs=preview_player).then(lambda: gr.update(visible=True), outputs=download_slice_btn)
+
+    create_vocals_loop_btn.click(fn=create_loop, inputs=[vocals_output, vocals_bar_times_state, vocals_loop_length], outputs=[vocals_loop_output, vocals_loop_download_btn])
+    create_drums_loop_btn.click(fn=create_loop, inputs=[drums_output, drums_bar_times_state, drums_loop_length], outputs=[drums_loop_output, drums_loop_download_btn])
+    create_bass_loop_btn.click(fn=create_loop, inputs=[bass_output, bass_bar_times_state, bass_loop_length], outputs=[bass_loop_output, bass_loop_download_btn])
+    create_other_loop_btn.click(fn=create_loop, inputs=[other_output, other_bar_times_state, other_loop_length], outputs=[other_loop_output, other_loop_download_btn])
+
+    download_slice_btn.click(fn=download_slice, inputs=preview_player, outputs=download_slice_btn)
+    cut_all_oneshots_btn.click(fn=cut_all_oneshots, inputs=[active_stem_state, onset_times_state], outputs=cut_oneshots_download_btn)
+
+
+demo.launch()