[音楽生成AI] ローカルで動かす音楽生成AI入門：MusicGenでゲームBGM/SEを作る」

# 1. 仮想環境を作成（推奨）
python -m venv musicgen-env
source musicgen-env/bin/activate  # Windowsは Scripts\activate

# 2. ライブラリインストール
pip install git+https://github.com/facebookresearch/audiocraft.git

# 3. テキストから音楽生成（英語）
python -m audiocraft.generate --prompt "ambient fantasy bgm with harp and strings"

C:\>cd C:\AI
C:\AI>mkdir musicgen
C:\AI>cd musicgen
C:\AI\musicgen>

C:\AI\musicgen>git clone https://github.com/facebookresearch/audiocraft
Cloning into 'audiocraft'...
remote: Enumerating objects: 2011, done.
remote: Counting objects: 100% (11/11), done.
remote: Compressing objects: 100% (5/5), done.
Receiving objects:  83% (1670/2011), 22.16 MiB | 21.69 MiB/sremote: Total 2011 (delta 8), reused 6 (delta 6), pack-reuseReceiving objec
Resolving deltas: 100% (1193/1193), done.

C:\AI\musicgen>cd audiocraft

C:\AI\musicgen\audiocraft>pip install -e .
Obtaining file:///C:/AI/musicgen/audiocraft
  Installing build dependencies ... done
  Checking if build backend supports build_editable ... done
  Getting requirements to build editable ... done
  Preparing editable metadata (pyproject.toml) ... done
Collecting av==11.0.0 (from audiocraft==1.4.0a2)
  Downloading av-11.0.0-cp310-cp310-win_amd64.whl.metadata (4.7 kB)
Requirement already satisfied: einops in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from audiocraft==1.4.0a2) (0.8.1)
Collecting flashy>=0.0.1 (from audiocraft==1.4.0a2)
  Downloading flashy-0.0.2.tar.gz (72 kB)
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
Collecting hydra-core>=1.1 (from audiocraft==1.4.0a2)
  Downloading hydra_core-1.3.2-py3-none-any.whl.metadata (5.5 kB)
Collecting hydra_colorlog (from audiocraft==1.4.0a2)
  Downloading hydra_colorlog-1.2.0-py3-none-any.whl.metadata (949 bytes)
Collecting julius (from audiocraft==1.4.0a2)
  Downloading julius-0.2.7.tar.gz (59 kB)
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
Collecting num2words (from audiocraft==1.4.0a2)
  Downloading num2words-0.5.14-py3-none-any.whl.metadata (13 kB)
Collecting numpy<2.0.0 (from audiocraft==1.4.0a2)
  Downloading numpy-1.26.4-cp310-cp310-win_amd64.whl.metadata (61 kB)
Collecting sentencepiece (from audiocraft==1.4.0a2)
  Using cached sentencepiece-0.2.0-cp310-cp310-win_amd64.whl.metadata (8.3 kB)
Collecting spacy==3.7.6 (from audiocraft==1.4.0a2)
  Downloading spacy-3.7.6-cp310-cp310-win_amd64.whl.metadata (27 kB)
Collecting torch==2.1.0 (from audiocraft==1.4.0a2)
  Downloading torch-2.1.0-cp310-cp310-win_amd64.whl.metadata (24 kB)
Collecting torchaudio<2.1.2,>=2.0.0 (from audiocraft==1.4.0a2)
  Downloading torchaudio-2.1.1-cp310-cp310-win_amd64.whl.metadata (6.4 kB)

（中略）

Successfully built audiocraft flashy demucs julius dora_search encodec docopt pesq treetable
Installing collected packages: sentencepiece, pesq, lameenc, docopt, cymem, antlr4-python3-runtime, wrapt, wasabi, treetable, threadpoolctl, spacy-loggers, spacy-legacy, retrying, pycparser, protobuf, platformdirs, omegaconf, numpy, num2words, murmurhash, msgpack, marisa-trie, llvmlite, lightning-utilities, lazy_loader, joblib, decorator, colorlog, cloudpickle, cloudpathlib, catalogue, av, audioread, torch, submitit, srsly, soxr, smart-open, scipy, preshed, pooch, numba, language-data, hydra-core, cffi, blis, xformers, torchvision, torchmetrics, torchdiffeq, torchdata, torchaudio, soundfile, scikit-learn, pystoi, langcodes, julius, hydra_colorlog, dora_search, confection, weasel, torchtext, thinc, openunmix, librosa, flashy, encodec, spacy, demucs, audiocraft
  Attempting uninstall: numpy
    Found existing installation: numpy 2.2.6
    Uninstalling numpy-2.2.6:
      Successfully uninstalled numpy-2.2.6
  Attempting uninstall: av
    Found existing installation: av 14.4.0
    Uninstalling av-14.4.0:
      Successfully uninstalled av-14.4.0
  Attempting uninstall: torch
    Found existing installation: torch 2.5.1+cu121
    Uninstalling torch-2.5.1+cu121:
      Successfully uninstalled torch-2.5.1+cu121
  Attempting uninstall: torchvision
    Found existing installation: torchvision 0.20.1+cu121
    Uninstalling torchvision-0.20.1+cu121:
      Successfully uninstalled torchvision-0.20.1+cu121
  Attempting uninstall: torchaudio
    Found existing installation: torchaudio 2.5.1+cu121
    Uninstalling torchaudio-2.5.1+cu121:
      Successfully uninstalled torchaudio-2.5.1+cu121
Successfully installed antlr4-python3-runtime-4.9.3 audiocraft-1.4.0a2 audioread-3.0.1 av-11.0.0 blis-0.7.11 catalogue-2.0.10 cffi-1.17.1 cloudpathlib-0.21.1 cloudpickle-3.1.1 colorlog-6.9.0 confection-0.1.5 cymem-2.0.11 decorator-5.2.1 demucs-4.0.1 docopt-0.6.2 dora_search-0.1.12 encodec-0.1.1 flashy-0.0.2 hydra-core-1.3.2 hydra_colorlog-1.2.0 joblib-1.5.1 julius-0.2.7 lameenc-1.8.1 langcodes-3.5.0 language-data-1.3.0 lazy_loader-0.4 librosa-0.11.0 lightning-utilities-0.14.3 llvmlite-0.44.0 marisa-trie-1.2.1 msgpack-1.1.1 murmurhash-1.0.13 num2words-0.5.14 numba-0.61.2 numpy-1.26.4 omegaconf-2.3.0 openunmix-1.3.0 pesq-0.0.4 platformdirs-4.3.8 pooch-1.8.2 preshed-3.0.10 protobuf-6.31.1 pycparser-2.22 pystoi-0.4.1 retrying-1.3.4 scikit-learn-1.7.0 scipy-1.15.3 sentencepiece-0.2.0 smart-open-7.1.0 soundfile-0.13.1 soxr-0.5.0.post1 spacy-3.7.6 spacy-legacy-3.0.12 spacy-loggers-1.0.5 srsly-2.5.1 submitit-1.5.3 thinc-8.2.5 threadpoolctl-3.6.0 torch-2.1.0 torchaudio-2.1.0 torchdata-0.7.0 torchdiffeq-0.2.5 torchmetrics-1.7.3 torchtext-0.16.0 torchvision-0.16.0 treetable-0.2.5 wasabi-1.1.3 weasel-0.4.1 wrapt-1.17.2 xformers-0.0.22.post7

C:\AI\musicgen\audiocraft>python -m audiocraft.models.download musicgen-medium
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.1.0+cu121 with CUDA 1201 (you have 2.1.0+cpu)
    Python  3.10.11 (you have 3.10.6)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details
C:\Users\user-name\AppData\Local\Programs\Python\Python310\python.exe: No module named audiocraft.models.download

from audiocraft.models import MusicGen

# 'small', 'medium', or 'large' のいずれか
model = MusicGen.get_pretrained('medium')

model.set_generation_params(duration=30)  # 秒数を指定
wav = model.generate(["lofi hip-hop beats"])
model.save(wav, 'output')

python generate_music.py

>python generate_music.py
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.1.0+cu121 with CUDA 1201 (you have 2.1.0+cpu)
    Python  3.10.11 (you have 3.10.6)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details
C:\AI\musicgen\audiocraft\audiocraft\models\musicgen.py:83: UserWarning: MusicGen pretrained model relying on deprecated checkpoint mapping. Please use full pre-trained id instead: facebook/musicgen-medium
  warnings.warn(
Xet Storage is enabled for this repo, but the 'hf_xet' package is not installed. Falling back to regular HTTP download. For better performance, install the package with: `pip install huggingface_hub[hf_xet]` or `pip install hf_xet`
state_dict.bin: 100%|█████████████████████████████████████████████████████████████| 3.68G/3.68G [00:44<00:00, 83.0MB/s]
C:\Users\user-name\AppData\Local\Programs\Python\Python310\lib\site-packages\huggingface_hub\file_download.py:143: UserWarning: `huggingface_hub` cache-system uses symlinks by default to efficiently store duplicated files but your machine does not support them in C:\Users\user-name\.cache\huggingface\hub\models--facebook--musicgen-medium. Caching files will still work but in a degraded version that might require more space on your disk. This warning can be disabled by setting the `HF_HUB_DISABLE_SYMLINKS_WARNING` environment variable. For more details, see https://huggingface.co/docs/huggingface_hub/how-to-cache#limitations.
To support symlinks on Windows, you either need to activate Developer Mode or to run Python as an administrator. In order to activate developer mode, see this article: https://docs.microsoft.com/en-us/windows/apps/get-started/enable-your-device-for-development
  warnings.warn(message)
spiece.model: 100%|█████████████████████████████████████████████████████████████████| 792k/792k [00:00<00:00, 1.69MB/s]
tokenizer.json: 100%|█████████████████████████████████████████████████████████████| 1.39M/1.39M [00:00<00:00, 2.14MB/s]
config.json: 100%|████████████████████████████████████████████████████████████████████████| 1.21k/1.21k [00:00<?, ?B/s]
model.safetensors: 100%|████████████████████████████████████████████████████████████| 892M/892M [00:12<00:00, 70.9MB/s]
Xet Storage is enabled for this repo, but the 'hf_xet' package is not installed. Falling back to regular HTTP download. For better performance, install the package with: `pip install huggingface_hub[hf_xet]` or `pip install hf_xet`
compression_state_dict.bin: 100%|███████████████████████████████████████████████████| 236M/236M [00:02<00:00, 89.5MB/s]
C:\Users\user-name\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\nn\utils\weight_norm.py:30: UserWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.
  warnings.warn("torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.")
Traceback (most recent call last):
  File "C:\AI\musicgen\audiocraft\generate_music.py", line 8, in <module>
    model.save(wav, 'output')
AttributeError: 'MusicGen' object has no attribute 'save'

WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions.

from audiocraft.models import MusicGen
import torchaudio

# モデルのロード
model = MusicGen.get_pretrained("facebook/musicgen-medium")
model.set_generation_params(duration=30)

# プロンプト設定
descriptions = ["lofi hip-hop beat with relaxing melody"]

# 音楽生成
wav = model.generate(descriptions)

# 出力保存（1曲目）
torchaudio.save("output.wav", wav[0].cpu(), sample_rate=32000)

from audiocraft.models import MusicGen
import torchaudio

# モデルのロード
model = MusicGen.get_pretrained("facebook/musicgen-medium")
model.set_generation_params(duration=30)

# プロンプト設定
descriptions = ["lofi hip-hop beat with relaxing melody"]

# 音楽生成
wav = model.generate(descriptions)

# 出力保存（1曲目）
torchaudio.save("output.wav", wav[0].cpu(), sample_rate=32000)

python generate_music.py

ffmpeg -i output.wav output.mp3

const bgm = new Audio('output.mp3');
bgm.loop = true;
bgm.play();

for i in range(5):
    wav = model.generate(["cyberpunk background music"])
    model.save(wav, f"cyberpunk_{i+1}.wav")

>python generate_music.py
WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.1.0+cu121 with CUDA 1201 (you have 2.1.0+cpu)
    Python  3.10.11 (you have 3.10.6)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
  Memory-efficient attention, SwiGLU, sparse and more won't be available.
  Set XFORMERS_MORE_DETAILS=1 for more details
C:\Users\user-name\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\nn\utils\weight_norm.py:30: UserWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.
  warnings.warn("torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.")

descriptions = ["ambient piano music with soft background texture"]

descriptions = ["simple 8-bit video game background music"]

model = MusicGen.get_pretrained('facebook/musicgen-medium')
model.set_generation_params(duration=10)

# GPUに転送
model = model.to('cuda')  # ← これが重要！

wav = model.generate_unconditional(1)

from audiocraft.models import MusicGen
import torch

# GPUが使えるか確認してデバイスを選択
device = 'cuda' if torch.cuda.is_available() else 'cpu'

# GPU対応でモデルを読み込む（これが正解！）
model = MusicGen.get_pretrained('facebook/musicgen-medium', device=device)

# 生成パラメータ（例：10秒）
model.set_generation_params(duration=10)

# 無条件で1曲生成
wav = model.generate_unconditional(1)

# 保存処理（必要に応じて）
model.save(wav, 'output')  # v1.2以降ならこれもOK（なければ自前で書く）

wav = model.generate_unconditional(1, device='cuda')

pip show audiocraft

pip uninstall torch torchaudio torchvision

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

>pip uninstall torch torchaudio torchvision
>pip uninstall torch torchaudio torchvision
Found existing installation: torch 2.1.0
Uninstalling torch-2.1.0:
  Would remove:
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\functorch\*
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\torch-2.1.0.dist-info\*
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\torch\*
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\torchgen\*
    c:\users\user-name\appdata\local\programs\python\python310\scripts\convert-caffe2-to-onnx.exe
    c:\users\user-name\appdata\local\programs\python\python310\scripts\convert-onnx-to-caffe2.exe
    c:\users\user-name\appdata\local\programs\python\python310\scripts\torchrun.exe
Proceed (Y/n)? y
  Successfully uninstalled torch-2.1.0
Found existing installation: torchaudio 2.1.0
Uninstalling torchaudio-2.1.0:
  Would remove:
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\torchaudio-2.1.0.dist-info\*
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\torchaudio\*
Proceed (Y/n)? y
  Successfully uninstalled torchaudio-2.1.0
Found existing installation: torchvision 0.16.0
Uninstalling torchvision-0.16.0:
  Would remove:
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\torchvision-0.16.0.dist-info\*
    c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages\torchvision\*
Proceed (Y/n)? y
  Successfully uninstalled torchvision-0.16.0

>pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
Looking in indexes: https://download.pytorch.org/whl/cu121
Collecting torch
  Using cached https://download.pytorch.org/whl/cu121/torch-2.5.1%2Bcu121-cp310-cp310-win_amd64.whl (2449.4 MB)
Collecting torchvision
  Using cached https://download.pytorch.org/whl/cu121/torchvision-0.20.1%2Bcu121-cp310-cp310-win_amd64.whl (6.1 MB)
Collecting torchaudio
  Using cached https://download.pytorch.org/whl/cu121/torchaudio-2.5.1%2Bcu121-cp310-cp310-win_amd64.whl (4.1 MB)
Requirement already satisfied: filelock in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torch) (3.18.0)
Requirement already satisfied: typing-extensions>=4.8.0 in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torch) (4.13.2)
Requirement already satisfied: networkx in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torch) (3.4.2)
Requirement already satisfied: jinja2 in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torch) (3.1.6)
Requirement already satisfied: fsspec in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torch) (2025.5.1)
Requirement already satisfied: sympy==1.13.1 in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torch) (1.13.1)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from sympy==1.13.1->torch) (1.3.0)
Requirement already satisfied: numpy in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torchvision) (1.26.4)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from torchvision) (10.1.0)
Requirement already satisfied: MarkupSafe>=2.0 in c:\users\user-name\appdata\local\programs\python\python310\lib\site-packages (from jinja2->torch) (3.0.2)
Installing collected packages: torch, torchvision, torchaudio
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
audiocraft 1.4.0a2 requires torch==2.1.0, but you have torch 2.5.1+cu121 which is incompatible.
audiocraft 1.4.0a2 requires torchaudio<2.1.2,>=2.0.0, but you have torchaudio 2.5.1+cu121 which is incompatible.
audiocraft 1.4.0a2 requires torchvision==0.16.0, but you have torchvision 0.20.1+cu121 which is incompatible.
torchdata 0.7.0 requires torch==2.1.0, but you have torch 2.5.1+cu121 which is incompatible.
torchtext 0.16.0 requires torch==2.1.0, but you have torch 2.5.1+cu121 which is incompatible.
xformers 0.0.22.post7 requires torch==2.1.0, but you have torch 2.5.1+cu121 which is incompatible.
Successfully installed torch-2.5.1+cu121 torchaudio-2.5.1+cu121 torchvision-0.20.1+cu121

pip uninstall audiocraft torch torchvision torchaudio torchtext torchdata xformers
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install git+https://github.com/facebookresearch/audiocraft.git

model.set_generation_params(
    duration=10,
    temperature=1.0,
    top_k=250,
    top_p=0.0,
    cfg_coef=3.0
)

wav = model.generate(
    ["epic orchestral soundtrack with brass and choir"],
    progress=True
)

model = MusicGen.get_pretrained("facebook/musicgen-medium")

prompts = [
    "upbeat electronic dance track with a strong beat",
    "lo-fi chill hip hop beat with vinyl crackle",
    "suspenseful orchestral soundtrack with strings"
]

model.set_generation_params(duration=12, temperature=1.1, top_k=250, cfg_coef=3.5)

for i, prompt in enumerate(prompts):
    wav = model.generate([prompt])
    torchaudio.save(f"out_{i+1}.wav", wav[0].cpu(), 32000)

import torch
from audiocraft.models import MusicGen
import torchaudio

prompts = [
    "upbeat electronic dance track with a strong beat",
    "lo-fi chill hip hop beat with vinyl crackle",
    "suspenseful orchestral soundtrack with strings"
]


print("✅ GPU available:", torch.cuda.is_available())
if torch.cuda.is_available():
    print("🚀 GPU:", torch.cuda.get_device_name(0))

# モデル読み込み
model = MusicGen.get_pretrained("facebook/musicgen-medium")
model.set_generation_params(duration=10)

# GPU転送（※ decoder は存在しないため除外）
model.lm.cuda()
model.compression_model.cuda()

model.set_generation_params(duration=12, temperature=1.1, top_k=250, cfg_coef=3.5)

for i, prompt in enumerate(prompts):
    wav = model.generate([prompt])
    torchaudio.save(f"out_{i+1}.wav", wav[0].cpu(), 32000)

upbeat electronic dance track with a strong beat

lo-fi chill hip hop beat with vinyl crackle

suspenseful orchestral soundtrack with strings

pip install pydub

from pydub import AudioSegment

sound = AudioSegment.from_wav("output.wav")

# クロスフェード処理
fade_len = 1000  # 1秒（ms）
loop = sound.fade_out(fade_len) + sound.fade_in(fade_len)

# ループを3回繰り返す
looped = loop * 3

# 書き出し
looped.export("output_looped.wav", format="wav")

loopy -i input.wav -o output.ogg --loop-start 10000 --loop-end 30000

ffmpeg -stream_loop 4 -i music.wav -i video.mp4 -shortest -c:v copy -c:a aac output.mp4