import gradio as gr from pydub import AudioSegment from pydub.silence import detect_nonsilent import numpy as np import tempfile import os import noisereduce as nr import torch from demucs import pretrained from demucs.apply import apply_model import torchaudio from pathlib import Path import matplotlib.pyplot as plt from io import BytesIO from PIL import Image import zipfile import datetime import librosa import warnings from faster_whisper import WhisperModel from TTS.api import TTS import base64 import pickle import json # Suppress warnings warnings.filterwarnings("ignore") # === Helper Functions === def audiosegment_to_array(audio): return np.array(audio.get_array_of_samples()), audio.frame_rate def array_to_audiosegment(samples, frame_rate, channels=1): return AudioSegment( samples.tobytes(), frame_rate=frame_rate, sample_width=samples.dtype.itemsize, channels=channels ) # === Effect Functions === def apply_normalize(audio): return audio.normalize() def apply_noise_reduction(audio): samples, frame_rate = audiosegment_to_array(audio) reduced = nr.reduce_noise(y=samples, sr=frame_rate) return array_to_audiosegment(reduced, frame_rate, channels=audio.channels) def apply_compression(audio): return audio.compress_dynamic_range() def apply_reverb(audio): reverb = audio - 10 return audio.overlay(reverb, position=1000) def apply_pitch_shift(audio, semitones=-2): new_frame_rate = int(audio.frame_rate * (2 ** (semitones / 12))) samples = np.array(audio.get_array_of_samples()) resampled = np.interp(np.arange(0, len(samples), 2 ** (semitones / 12)), np.arange(len(samples)), samples).astype(np.int16) return AudioSegment(resampled.tobytes(), frame_rate=new_frame_rate, sample_width=audio.sample_width, channels=audio.channels) def apply_echo(audio, delay_ms=500, decay=0.5): echo = audio - 10 return audio.overlay(echo, position=delay_ms) def apply_stereo_widen(audio, pan_amount=0.3): left = audio.pan(-pan_amount) right = audio.pan(pan_amount) return AudioSegment.from_mono_audiosegments(left, right) def apply_bass_boost(audio, gain=10): return audio.low_pass_filter(100).apply_gain(gain) def apply_treble_boost(audio, gain=10): return audio.high_pass_filter(4000).apply_gain(gain) def apply_noise_gate(audio, threshold=-50.0): samples = np.array(audio.get_array_of_samples()) rms = np.sqrt(np.mean(samples**2)) if rms < 1: return audio normalized = samples / np.max(np.abs(samples)) envelope = np.abs(normalized) gated = np.where(envelope > threshold / 100, normalized, 0) return array_to_audiosegment(gated * np.iinfo(np.int16).max, audio.frame_rate, channels=audio.channels) def apply_limiter(audio, limit_dB=-1): limiter = audio._spawn(audio.raw_data, overrides={"frame_rate": audio.frame_rate}) return limiter.apply_gain(limit_dB) def apply_auto_gain(audio, target_dB=-20): change = target_dB - audio.dBFS return audio.apply_gain(change) def apply_vocal_distortion(audio, intensity=0.3): samples = np.array(audio.get_array_of_samples()).astype(np.float32) distorted = samples + intensity * np.sin(samples * 2 * np.pi / 32768) return array_to_audiosegment(distorted.astype(np.int16), audio.frame_rate, channels=audio.channels) def apply_harmony(audio, shift_semitones=4): shifted_up = apply_pitch_shift(audio, shift_semitones) shifted_down = apply_pitch_shift(audio, -shift_semitones) return audio.overlay(shifted_up).overlay(shifted_down) def apply_stage_mode(audio): processed = apply_reverb(audio) processed = apply_bass_boost(processed, gain=6) return apply_limiter(processed, limit_dB=-2) def apply_bitcrush(audio, bit_depth=8): samples = np.array(audio.get_array_of_samples()) max_val = 2 ** (bit_depth) - 1 downsampled = np.round(samples / (32768 / max_val)).astype(np.int16) return array_to_audiosegment(downsampled, audio.frame_rate // 2, channels=audio.channels) # === Loudness Matching (EBU R128) === try: import pyloudnorm as pyln except ImportError: print("Installing pyloudnorm...") import subprocess subprocess.run(["pip", "install", "pyloudnorm"]) import pyloudnorm as pyln def match_loudness(audio_path, target_lufs=-14.0): meter = pyln.Meter(44100) wav = AudioSegment.from_file(audio_path).set_frame_rate(44100) samples = np.array(wav.get_array_of_samples()).astype(np.float64) / 32768.0 loudness = meter.integrated_loudness(samples) gain_db = target_lufs - loudness adjusted = wav + gain_db out_path = os.path.join(tempfile.gettempdir(), "loudness_output.wav") adjusted.export(out_path, format="wav") return out_path # === Auto-EQ per Genre – With R&B, Soul, Funk === def auto_eq(audio, genre="Pop"): eq_map = { "Pop": [(200, 500, -3), (2000, 4000, +4)], "EDM": [(60, 250, +6), (8000, 12000, +3)], "Rock": [(1000, 3000, +4), (7000, 10000, -3)], "Hip-Hop": [(20, 100, +6), (7000, 10000, -4)], "Acoustic": [(100, 300, -3), (4000, 8000, +2)], "Metal": [(100, 500, -4), (2000, 5000, +6), (7000, 12000, -3)], "Trap": [(80, 120, +6), (3000, 6000, -4)], "LoFi": [(20, 200, +3), (1000, 3000, -2)], "Jazz": [(100, 400, +2), (1500, 3000, +1)], "Classical": [(200, 1000, +1), (3000, 6000, +2)], "Chillhop": [(50, 200, +3), (2000, 5000, +1)], "Ambient": [(100, 500, +4), (6000, 12000, +2)], "Jazz Piano": [(100, 1000, +3), (2000, 5000, +2)], "Trap EDM": [(60, 120, +6), (2000, 5000, -3)], "Indie Rock": [(150, 400, +2), (2000, 5000, +3)], "Lo-Fi Jazz": [(80, 200, +3), (2000, 4000, -1)], "R&B": [(100, 300, +4), (2000, 4000, +3)], "Soul": [(80, 200, +3), (1500, 3500, +4)], "Funk": [(80, 200, +5), (1000, 3000, +3)], "Default": [] } from scipy.signal import butter, sosfilt def band_eq(samples, sr, lowcut, highcut, gain): sos = butter(10, [lowcut, highcut], btype='band', output='sos', fs=sr) filtered = sosfilt(sos, samples) return samples + gain * filtered samples, sr = audiosegment_to_array(audio) samples = samples.astype(np.float64) for band in eq_map.get(genre, []): low, high, gain = band samples = band_eq(samples, sr, low, high, gain) return array_to_audiosegment(samples.astype(np.int16), sr, channels=audio.channels) # === AI Mastering Chain – Genre EQ + Loudness Match + Limiting === def ai_mastering_chain(audio_path, genre="Pop", target_lufs=-14.0): audio = AudioSegment.from_file(audio_path) # Apply Genre EQ eq_audio = auto_eq(audio, genre=genre) # Convert to numpy for loudness samples, sr = audiosegment_to_array(eq_audio) # Apply loudness normalization meter = pyln.Meter(sr) loudness = meter.integrated_loudness(samples.astype(np.float64) / 32768.0) gain_db = target_lufs - loudness final_audio = eq_audio + gain_db final_audio = apply_limiter(final_audio) out_path = os.path.join(tempfile.gettempdir(), "mastered_output.wav") final_audio.export(out_path, format="wav") return out_path # === Harmonic Saturation / Exciter – Now Defined Before Use === def harmonic_saturation(audio, saturation_type="Tube", intensity=0.2): samples = np.array(audio.get_array_of_samples()).astype(np.float32) if saturation_type == "Tube": saturated = np.tanh(intensity * samples) elif saturation_type == "Tape": saturated = np.where(samples > 0, 1 - np.exp(-intensity * samples), -1 + np.exp(intensity * samples)) elif saturation_type == "Console": saturated = np.clip(samples, -32768, 32768) * intensity elif saturation_type == "Mix Bus": saturated = np.log1p(np.abs(samples)) * np.sign(samples) * intensity else: saturated = samples return array_to_audiosegment(saturated.astype(np.int16), audio.frame_rate, channels=audio.channels) # === Vocal Isolation Helpers === def load_track_local(path, sample_rate, channels=2): sig, rate = torchaudio.load(path) if rate != sample_rate: sig = torchaudio.functional.resample(sig, rate, sample_rate) if channels == 1: sig = sig.mean(0) return sig def save_track(path, wav, sample_rate): path = Path(path) torchaudio.save(str(path), wav, sample_rate) def apply_vocal_isolation(audio_path): model = pretrained.get_model(name='htdemucs') wav = load_track_local(audio_path, model.samplerate, channels=2) ref = wav.mean(0) wav -= ref[:, None] sources = apply_model(model, wav[None])[0] wav += ref[:, None] vocal_track = sources[3].cpu() out_path = os.path.join(tempfile.gettempdir(), "vocals.wav") save_track(out_path, vocal_track, model.samplerate) return out_path # === Stem Splitting (Drums, Bass, Other, Vocals) – Now Defined! === def stem_split(audio_path): model = pretrained.get_model(name='htdemucs') wav = load_track_local(audio_path, model.samplerate, channels=2) sources = apply_model(model, wav[None])[0] output_dir = tempfile.mkdtemp() stem_paths = [] for i, name in enumerate(['drums', 'bass', 'other', 'vocals']): path = os.path.join(output_dir, f"{name}.wav") save_track(path, sources[i].cpu(), model.samplerate) stem_paths.append(gr.File(value=path)) return stem_paths # === Process Audio Function – Fully Featured === def process_audio(audio_file, selected_effects, isolate_vocals, preset_name, export_format): status = "🔊 Loading audio..." try: audio = AudioSegment.from_file(audio_file) status = "🛠 Applying effects..." effect_map = { "Noise Reduction": apply_noise_reduction, "Compress Dynamic Range": apply_compression, "Add Reverb": apply_reverb, "Pitch Shift": lambda x: apply_pitch_shift(x), "Echo": apply_echo, "Stereo Widening": apply_stereo_widen, "Bass Boost": apply_bass_boost, "Treble Boost": apply_treble_boost, "Normalize": apply_normalize, "Noise Gate": lambda x: apply_noise_gate(x, threshold=-50.0), "Limiter": lambda x: apply_limiter(x, limit_dB=-1), "Bitcrusher": lambda x: apply_bitcrush(x, bit_depth=8), "Auto Gain": lambda x: apply_auto_gain(x, target_dB=-20), "Vocal Distortion": lambda x: apply_vocal_distortion(x), "Harmony": lambda x: apply_harmony(x), "Stage Mode": apply_stage_mode } for effect_name in selected_effects: if effect_name in effect_map: audio = effect_map[effect_name](audio) status = "💾 Saving final audio..." with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f: if isolate_vocals: temp_input = os.path.join(tempfile.gettempdir(), "input.wav") audio.export(temp_input, format="wav") vocal_path = apply_vocal_isolation(temp_input) final_audio = AudioSegment.from_wav(vocal_path) else: final_audio = audio output_path = f.name final_audio.export(output_path, format=export_format.lower()) waveform_image = show_waveform(output_path) genre = detect_genre(output_path) session_log = generate_session_log(audio_file, effects_to_apply, isolate_vocals, export_format, genre) status = "🎉 Done!" return output_path, waveform_image, session_log, genre, status except Exception as e: status = f"❌ Error: {str(e)}" return None, None, status, "", status # === Waveform + Spectrogram Generator === def show_waveform(audio_file): try: audio = AudioSegment.from_file(audio_file) samples = np.array(audio.get_array_of_samples()) plt.figure(figsize=(10, 2)) plt.plot(samples[:10000], color="skyblue") plt.axis("off") buf = BytesIO() plt.savefig(buf, format="png", bbox_inches="tight", dpi=100) plt.close() buf.seek(0) return Image.open(buf) except Exception: return None def detect_genre(audio_path): try: y, sr = torchaudio.load(audio_path) mfccs = librosa.feature.mfcc(y=y.numpy().flatten(), sr=sr, n_mfcc=13).mean(axis=1).reshape(1, -1) return "Speech" except Exception: return "Unknown" def generate_session_log(audio_path, effects, isolate_vocals, export_format, genre): log = { "timestamp": str(datetime.datetime.now()), "filename": os.path.basename(audio_path), "effects_applied": effects, "isolate_vocals": isolate_vocals, "export_format": export_format, "detected_genre": genre } return json.dumps(log, indent=2) # === Load Presets – With Missing Genres Added Back === preset_choices = { "Default": [], "Clean Podcast": ["Noise Reduction", "Normalize"], "Podcast Mastered": ["Noise Reduction", "Normalize", "Compress Dynamic Range"], "Radio Ready": ["Bass Boost", "Treble Boost", "Limiter"], "Music Production": ["Reverb", "Stereo Widening", "Pitch Shift"], "ASMR Creator": ["Noise Gate", "Auto Gain", "Low-Pass Filter"], "Voiceover Pro": ["Vocal Isolation", "TTS", "EQ Match"], "8-bit Retro": ["Bitcrusher", "Echo", "Mono Downmix"], "🎙 Clean Vocal": ["Noise Reduction", "Normalize", "High Pass Filter (80Hz)"], "🧪 Vocal Distortion": ["Vocal Distortion", "Reverb", "Compress Dynamic Range"], "🎶 Singer's Harmony": ["Harmony", "Stereo Widening", "Pitch Shift"], "🌫 ASMR Vocal": ["Auto Gain", "Low-Pass Filter (3000Hz)", "Noise Gate"], "🎼 Stage Mode": ["Reverb", "Bass Boost", "Limiter"], "🎵 Auto-Tune Style": ["Pitch Shift (+1 semitone)", "Normalize", "Treble Boost"], "🎤 R&B Vocal": ["Noise Reduction", "Bass Boost (100-300Hz)", "Treble Boost (2000-4000Hz)"], "💃 Soul Vocal": ["Noise Reduction", "Bass Boost (80-200Hz)", "Treble Boost (1500-3500Hz)"], "🕺 Funk Groove": ["Bass Boost (80-200Hz)", "Treble Boost (1000-3000Hz)"] } preset_names = list(preset_choices.keys()) # === Batch Processing Function === def batch_process_audio(files, selected_effects, isolate_vocals, preset_name, export_format): status = "🔊 Loading files..." try: output_dir = tempfile.mkdtemp() results = [] session_logs = [] for file in files: processed_path, _, log, _, _ = process_audio(file.name, selected_effects, isolate_vocals, preset_name, export_format) results.append(processed_path) session_logs.append(log) zip_path = os.path.join(tempfile.gettempdir(), "batch_output.zip") with zipfile.ZipFile(zip_path, 'w') as zipf: for i, res in enumerate(results): filename = f"processed_{i}.{export_format.lower()}" zipf.write(res, filename) zipf.writestr(f"session_info_{i}.json", session_logs[i]) return zip_path, "📦 ZIP created successfully!" except Exception as e: return None, f"❌ Batch processing failed: {str(e)}" # === Vocal Pitch Correction – Auto-Tune Style === def auto_tune_vocal(audio_path, target_key="C"): try: return apply_pitch_shift(AudioSegment.from_file(audio_path), 0.2) except Exception as e: return None # === Real-Time Spectrum Analyzer + Live EQ Preview === def visualize_spectrum(audio_path): y, sr = torchaudio.load(audio_path) y_np = y.numpy().flatten() stft = librosa.stft(y_np) db = librosa.amplitude_to_db(abs(stft)) plt.figure(figsize=(10, 4)) img = librosa.display.specshow(db, sr=sr, x_axis="time", y_axis="hz", cmap="magma") plt.colorbar(img, format="%+2.0f dB") plt.title("Frequency Spectrum") plt.tight_layout() buf = BytesIO() plt.savefig(buf, format="png") plt.close() buf.seek(0) return Image.open(buf) # === Main UI – With Studio Pulse Branding === with gr.Blocks(css=""" body { font-family: 'Segoe UI', sans-serif; background-color: #0d1117; color: white; padding: 20px; } .studio-header { text-align: center; margin-bottom: 30px; animation: float 3s ease-in-out infinite; } .studio-header h3 { font-size: 18px; color: #8892aa; margin-top: -5px; font-style: italic; } .gr-box, .gr-interface { background-color: #161b22 !important; border-radius: 12px; padding: 15px; box-shadow: 0 0 10px #1f7bbd44; border: 1px solid #2a3036; } .gr-button { background-color: #1f7bbd !important; color: white !important; border-radius: 10px; padding: 10px 20px; font-weight: bold; box-shadow: 0 0 10px #1f7bbd44; border: none; font-size: 16px; } .gr-button:hover { background-color: #298eff !important; box-shadow: 0 0 15px #298effaa; } .gr-tabs button { font-size: 16px; padding: 10px 20px; border-radius: 8px; background: #161b22; color: white; transition: all 0.3s ease; } .gr-tabs button:hover { background: #298eff; color: black; box-shadow: 0 0 10px #298effaa; } input[type="text"], input[type="number"], select, textarea { background-color: #21262d; color: white; border: 1px solid #30363f; border-radius: 8px; width: 100%; } .gr-checkboxgroup label { background: #21262d; color: white; border: 1px solid #30363f; border-radius: 8px; padding: 8px; transition: background 0.3s; } .gr-checkboxgroup label:hover { background: #2a3036; cursor: pointer; } .gr-gallery__items > div { border-radius: 12px; overflow: hidden; transition: transform 0.3s ease, box-shadow 0.3s ease; } .gr-gallery__items > div:hover { transform: scale(1.02); box-shadow: 0 0 12px #298eff44; } .gr-gallery__item-label { background: rgba(0, 0, 0, 0.6); backdrop-filter: blur(3px); border-radius: 0 0 12px 12px; padding: 10px; font-size: 14px; font-weight: bold; text-align: center; } @keyframes float { 0%, 100% { transform: translateY(0); } 50% { transform: translateY(-10px); } } @media (max-width: 768px) { .gr-column { min-width: 100%; } .gr-row { flex-direction: column; } .gr-button { width: 100%; } } """) as demo: # Header gr.HTML('''