music_note Génération audio avec Stable Audio

mkdir stable-audio-tuto
cd stable-audio-tuto

python -m venv venv
venv\Scripts\activate            # Windows
source venv/bin/activate         # macOS / Linux

# PyTorch avec CUDA (adapter la version à ton setup sur pytorch.org)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121

# stable-audio-tools : la bibliothèque officielle de Stability AI
pip install stable-audio-tools

# soundfile : pour sauvegarder l'audio en WAV
pip install soundfile huggingface_hub

import torch
import soundfile as sf
from stable_audio_tools import get_pretrained_model
from stable_audio_tools.inference.generation import generate_diffusion_cond

# ── Charger le modèle ─────────────────────────────────────────────────────────
# La première exécution télécharge ~3.5 Go depuis HuggingFace.
# Les fois suivantes, il est en cache local (~/.cache/huggingface).

print("Chargement du modèle Stable Audio Open...")

device = "cuda" if torch.cuda.is_available() else "cpu"
model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
model = model.to(device)

sample_rate = model_config["sample_rate"]   # 44100 Hz
sample_size = model_config["sample_size"]   # taille max du sample

print(f"Modèle chargé sur : {device}")

# ── Prompt ────────────────────────────────────────────────────────────────────
# Décris : style, instruments, BPM, ambiance.
# Les prompts en anglais donnent de meilleurs résultats.
prompt = (
    "atmospheric dark ambient music, slow evolving pads, "
    "deep bass drone, cinematic tension, no drums, 130 BPM"
)
negative_prompt = "low quality, noise, distortion, vocals, speech"

duree_secondes = 30  # durée souhaitée (max ~47s)

print(f"Génération de {duree_secondes} secondes d'audio...")

# ── Générer ───────────────────────────────────────────────────────────────────
conditioning = [
    {
        "prompt":         prompt,
        "seconds_start":  0,
        "seconds_total":  duree_secondes,
    }
]
negative_conditioning = [
    {
        "prompt":         negative_prompt,
        "seconds_start":  0,
        "seconds_total":  duree_secondes,
    }
]

with torch.no_grad():
    audio = generate_diffusion_cond(
        model,
        steps=100,                      # étapes de diffusion (50-200)
        cfg_scale=7.0,                  # guidance scale
        conditioning=conditioning,
        negative_conditioning=negative_conditioning,
        batch_size=1,
        sample_size=sample_size,
        sigma_min=0.3,
        sigma_max=500,
        sampler_type="dpmpp-3m-sde",    # algorithme d'échantillonnage
        device=device,
        seed=42,
    )

# ── Sauvegarder ───────────────────────────────────────────────────────────────
# audio shape : [batch, channels, samples] → on prend [0] et on transpose
audio_numpy = audio[0].cpu().float().numpy().T       # [samples, channels]
n_samples   = int(duree_secondes * sample_rate)
audio_numpy = audio_numpy[:n_samples]                # rogner à la durée voulue

sf.write("output_audio.wav", audio_numpy, sample_rate)
print("✓ Audio sauvegardé : output_audio.wav")

python generate_audio.py

# Loop de batterie hip-hop
prompt = "punchy drum loop, hip hop, 90 BPM, crispy snare, deep kick, vinyl crackle"

# Ambiance nature
prompt = "forest ambience, birds chirping, light wind through leaves, distant stream, peaceful, morning"

# Techno
prompt = "techno music loop, 140 BPM, driving kick drum, acid bassline, hypnotic, dark warehouse"

# Cinématique
prompt = "epic orchestral music, strings, brass, choir, building tension, cinematic"

# Effet sonore sci-fi
prompt = "sci-fi interface beep, futuristic UI sound, short, clean, sterile"

# Jazz lo-fi
prompt = "lo-fi jazz, piano, double bass, brushed drums, 80 BPM, warm, nostalgic, vinyl noise"

# Guitare modale (tu vois ce que c'est...)
prompt = "modal guitar improvisation, open tuning, atmospheric, reverb, slow, meditative, no drums"

import torch
import soundfile as sf
from stable_audio_tools import get_pretrained_model
from stable_audio_tools.inference.generation import generate_diffusion_cond

device = "cuda" if torch.cuda.is_available() else "cpu"
model, model_config = get_pretrained_model("stabilityai/stable-audio-open-1.0")
model = model.to(device)

sample_rate = model_config["sample_rate"]
sample_size = model_config["sample_size"]

prompt = "atmospheric ambient drone, evolving textures, deep space, no rhythm"
duree  = 20
seeds  = [42, 123, 456, 789]

for seed in seeds:
    print(f"Génération seed {seed}...")
    conditioning          = [{"prompt": prompt, "seconds_start": 0, "seconds_total": duree}]
    negative_conditioning = [{"prompt": "low quality, noise", "seconds_start": 0, "seconds_total": duree}]

    with torch.no_grad():
        audio = generate_diffusion_cond(
            model, steps=80, cfg_scale=7.0,
            conditioning=conditioning,
            negative_conditioning=negative_conditioning,
            batch_size=1, sample_size=sample_size,
            sigma_min=0.3, sigma_max=500,
            sampler_type="dpmpp-3m-sde", device=device, seed=seed,
        )

    audio_np = audio[0].cpu().float().numpy().T
    audio_np = audio_np[:int(duree * sample_rate)]
    sf.write(f"variation_seed_{seed}.wav", audio_np, sample_rate)
    print(f"  → variation_seed_{seed}.wav sauvegardé")

print("Terminé !")