import os import json import time import re import tempfile from io import BytesIO import threading import gradio as gr from dotenv import load_dotenv from google.oauth2 import service_account from googleapiclient.discovery import build from googleapiclient.http import MediaIoBaseDownload import numpy as np import librosa import soundfile as sf # === Load Google Drive service account === load_dotenv() SERVICE_ACCOUNT_JSON = os.environ.get("SERVICE_ACCOUNT_JSON") if not SERVICE_ACCOUNT_JSON: raise RuntimeError("Missing SERVICE_ACCOUNT_JSON") creds_dict = json.loads(SERVICE_ACCOUNT_JSON) creds = service_account.Credentials.from_service_account_info( creds_dict, scopes=["https://www.googleapis.com/auth/drive.readonly"] ) drive_service = build("drive", "v3", credentials=creds) # === Google Drive audio file indexing === ROOT_FOLDER_ID = "1l7-LY9Vf9lDs8ltjDqckxrqUl_hj3W4d" def _list_audio_files(folder_id, path=""): results = drive_service.files().list( q=f"'{folder_id}' in parents and trashed = false", fields="files(id, name, mimeType)" ).execute() file_map = {} for item in results.get("files", []): name = item["name"] full_path = f"{path}/{name}".strip("/") if item["mimeType"] == "application/vnd.google-apps.folder": file_map.update(_list_audio_files(item["id"], full_path)) elif item["mimeType"].startswith("audio/"): file_map[full_path] = item["id"] return file_map audio_library = _list_audio_files(ROOT_FOLDER_ID) # === Parse metadata from filenames === pattern = r'_T(\d+)_|0*(\d+)\.|_(\d)_' titles = {} for filename, fid in audio_library.items(): if not filename.lower().endswith(('.flac', '.wav', '.aac')): continue parts = filename.split('/') if len(parts) < 3: continue match = re.search(pattern, filename) if match: num_str = next(g for g in match.groups() if g is not None) num = int(num_str) else: print(f"⚠️ Could not parse track number: {filename}") continue if '._instrumental_ASHQ' in filename: stem = 'instrumental_ASHQ' elif '._instrumental_AS' in filename or '._instrumentalAS' in filename: stem = 'instrumental_AS' elif '._instrumental' in filename: stem = 'instrumental' elif '._vocals_ASHQ' in filename: stem = 'vocals_ASHQ' elif '._vocals_AS' in filename or '._vocalsAS' in filename: stem = 'vocals_AS' elif '._vocals' in filename: stem = 'vocals' else: stem = 'orig' titles[(parts[1], parts[2], num, stem)] = fid # === Audio choices === format_choices = ['AAC-96kbps', 'AAC-256kbps', 'FLACs'] stem_choices = [ 'orig', 'instrumental', 'instrumental_AS', 'instrumental_ASHQ', 'vocals', 'vocals_AS', 'vocals_ASHQ' ] # === In-memory storage === audio_tracks = { (f, s): gr.Audio(label=f"{f} / {s}", type="numpy", visible=False) for f in format_choices for s in stem_choices } downloaded_audio_flacs = {} # (format, stem) -> BytesIO FLAC # === Download helper === def _download_audio(library, frmt, track, stem): file_id = titles.get((library, frmt, track, stem)) if not file_id: raise ValueError(f"Missing file: {library}, {frmt}, {track}, {stem}") buffer = BytesIO() request = drive_service.files().get_media(fileId=file_id) downloader = MediaIoBaseDownload(buffer, request) done = False while not done: status, done = downloader.next_chunk() print(f"⬇️ Downloaded {int(status.progress() * 100)}%") buffer.seek(0) with tempfile.NamedTemporaryFile(suffix=".audio", delete=False) as tmp: tmp.write(buffer.read()) tmp_path = tmp.name y, sr = librosa.load(tmp_path, sr=None, mono=False) # keep stereo os.remove(tmp_path) if y.ndim == 2: y = y.T # Convert from (channels, samples) → (samples, channels) flac_buf = BytesIO() sf.write(flac_buf, y, sr, format="FLAC") flac_buf.seek(0) downloaded_audio_flacs[(frmt, stem)] = flac_buf # === Load UI handler === def _load(library, track, formats, stems): selected_keys = set((f, s) for f in formats for s in stems) outputs = [] status_lines = [] for (f, s), audio in audio_tracks.items(): if (f, s) in selected_keys: try: _download_audio(library, f, track, s) flac_buf = downloaded_audio_flacs[(f, s)] flac_buf.seek(0) y, sr = sf.read(flac_buf, dtype='float32') outputs.append(gr.update(value=(sr, y), visible=True, label=f"{f} / {s}")) # status_lines.append(f"✅ Loaded {f} / {s}") except Exception as e: outputs.append(gr.update(value=None, visible=False)) status_lines.append(f"❌ Failed {f} / {s}: {str(e)}") else: outputs.append(gr.update(value=None, visible=False)) available_keys = [f"{f} / {s}" for (f, s) in downloaded_audio_flacs.keys()] instrumental_key = None for k in available_keys: if 'instrumental' in k: instrumental_key = k break vocals_key = None for k in available_keys: if 'vocals' in k: vocals_key = k break mix1_update = gr.update(choices=available_keys, value=instrumental_key) mix2_update = gr.update(choices=available_keys, value=vocals_key) status_text_output = "\n".join(status_lines) if status_lines else "" return outputs + [gr.update(value=status_text_output), mix1_update, mix2_update] # === Mix handler === def _mix_tracks(track1_key, track2_key, level, subtract_mix): if not track1_key or not track2_key: raise gr.Error("Please select both tracks.") key1 = tuple(track1_key.split(" / ")) key2 = tuple(track2_key.split(" / ")) buf1 = downloaded_audio_flacs.get(key1) buf2 = downloaded_audio_flacs.get(key2) if buf1 is None or buf2 is None: raise gr.Error("Selected tracks must be loaded first.") buf1.seek(0) buf2.seek(0) y1, sr1 = sf.read(buf1, dtype='float32') y2, sr2 = sf.read(buf2, dtype='float32') if sr1 != sr2: raise gr.Error("Sampling rates do not match.") min_len = min(len(y1), len(y2)) y1 = y1[:min_len] y2 = y2[:min_len] y1_level = y2_level = 1. if level < .5: y2_level = 2 * level else: y1_level = 2 - (2 * level) if subtract_mix: mixed = y1 - level * y2 else: # mixed = np.clip(y1_level * y1 + y2_level * y2, -1.0, 1.0) mixed = np.clip((1-level) * y1 + level * y2, -1.0, 1.0) return (sr1, mixed) # === UI === with gr.Blocks() as demo: gr.Markdown("# 🎵 Stems Evaluation for Realize Music") gr.Markdown( "❗️ **Warning**: This app is not compatible with **Safari** or mobile browsers! Please use a desktop browser." ) with gr.Tab("Stem Selection"): with gr.Row(): library_dropdown = gr.Dropdown( choices=["UM-00600753511770", "Virgin-1405031933546", "Warner-5034644802213"], label="Select Library", value="Warner-5034644802213" ) track_dropdown = gr.Dropdown(choices=[1, 2, 3, 4], label="Track Number") load_button = gr.Button("Load") with gr.Row(): selected_formats = gr.CheckboxGroup( choices=format_choices, label="Audio Input Formats", value=["AAC-96kbps"] ) selected_stems = gr.CheckboxGroup( choices=stem_choices, label="Stems to Retrieve", value=["orig", "instrumental_AS", "vocals_AS"] ) status_text = gr.Textbox(label="Status", interactive=False) gr.Markdown("### 🎧 Loaded Audio Grid") with gr.Column(): for stem in stem_choices: with gr.Row(): for frmt in format_choices: audio_tracks[(frmt, stem)].render() # === Mix Tab === with gr.Tab("🔀 Mix Tracks"): gr.Markdown("### Mix any two loaded audio tracks") gr.Markdown("For **subtracting**, orig-vocals=1.0 should be instrumental and orig-instrumental=1.0 should be vocals. And subtracting low-res from high-quality should show the residual difference. But note, playback is normalized and artifacts might sound much greater than they are.") gr.Markdown("(After changing the mix level, you have to select `Mix Tracks` again.)") with gr.Row(): mix1_dropdown = gr.Dropdown(label="Track 1") mix2_dropdown = gr.Dropdown(label="Track 2") mix_slider = gr.Slider(minimum=0, maximum=1, value=0.5, step=0.01, label="Mix (Trk1<->Trk2)") mix_subtract_chk = gr.Checkbox(label="subtract (Trk1-Trk2)", value=False) mix_button = gr.Button("Mix Tracks") mix_output = gr.Audio(type="numpy", label="Mixed Output") mix_button.click( fn=_mix_tracks, inputs=[mix1_dropdown, mix2_dropdown, mix_slider, mix_subtract_chk], outputs=mix_output ) load_button.click( fn=_load, inputs=[library_dropdown, track_dropdown, selected_formats, selected_stems], outputs=list(audio_tracks.values()) + [status_text, mix1_dropdown, mix2_dropdown] ) if __name__ == "__main__": demo.launch()