Local AI BGM and sound effects for Claude Code and any MCP client. Stable Audio 3 on Apple Silicon (MLX) and Windows NVIDIA GPUs (PyTorch). No cloud, no API key.
Python apple-siliconaudio-generationclaude-codecudagame-audiomcpmcp-servermlxmusic-generationpytorchsound-effectsstable-audiowindows
import shutil
import subprocess
import sys
import numpy as np
import pytest
from audio_studio_mcp import post
from conftest import sine
needs_ffmpeg = pytest.mark.skipif(shutil.which("ffmpeg") is None, reason="ffmpeg not installed")
def test_peak_ceiling_only_turns_down():
quiet = sine(amp=0.1)
out, gain = post.peak_ceiling(quiet, -1.0)
assert gain == 0.0 and out is quiet
loud = sine(amp=1.5)
out, gain = post.peak_ceiling(loud, -1.0)
assert gain < 0
assert post.stats(out)["peak_dbfs"] == pytest.approx(-1.0, abs=0.01)
def test_peak_normalize_boosts():
out, gain = post.peak_normalize(sine(amp=0.1), -1.0)
assert gain > 0
assert post.stats(out)["peak_dbfs"] == pytest.approx(-1.0, abs=0.01)
def test_trim():
a = sine(2.0, sr=1000)
assert post.trim(a, 1000, 0.5, 1.5).shape == (2, 1000)
with pytest.raises(ValueError):
post.trim(a, 1000, 1.5, 1.0)
def test_trim_silence():
sr = 44100
a = np.concatenate([np.zeros((2, sr)), sine(1.0, sr), np.zeros((2, sr))], axis=1).astype(np.float32)
out = post.trim_silence(a, sr, -50)
assert 1.0 <= out.shape[-1] / sr < 1.1
def test_fade_edges_are_silent():
out = post.fade(sine(1.0), 44100, 100, 100)
assert abs(out[:, 0]).max() == 0 and abs(out[:, -1]).max() < 1e-3
def test_make_loop_is_continuous():
sr = 1000
ramp = np.tile(np.linspace(0, 1, 3 * sr, dtype=np.float32), (2, 1))
out = post.make_loop(ramp, sr, 500)
assert out.shape[-1] == 3 * sr - 500
# the last sample flows into the first: the start equals the original tail
assert out[0, 0] == pytest.approx(ramp[0, -500], abs=1e-3)
with pytest.raises(ValueError):
post.make_loop(ramp, sr, 2000)
def test_to_mono():
assert post.to_mono(sine()).shape[0] == 1
@pytest.mark.parametrize("fmt,ext", [("wav", ".wav"), ("flac", ".flac"), ("ogg", ".ogg")])
def test_write_read_roundtrip(tmp_path, fmt, ext):
path = tmp_path / f"x{ext}"
post.write(path, sine(0.5), 44100, fmt)
audio, sr = post.read(path)
assert sr == 44100 and audio.shape[0] == 2
assert abs(audio.shape[-1] - 22050) < 2000 # vorbis pads a little
def test_long_ogg_write_does_not_crash_worker_thread(tmp_path):
# A long clip handed to libsndfile's Vorbis encoder in one call overflowed
# the 1 MB Windows thread stack and killed the process (#10). Run it in a
# child process so a crash fails this test instead of pytest.
path = tmp_path / "long.ogg"
code = (
"import sys, threading, numpy as np\n"
"from audio_studio_mcp import post\n"
"a = (np.random.default_rng(0).random((2, 48000 * 60), dtype=np.float32) - 0.5) * 0.5\n"
"t = threading.Thread(target=post.write, args=(sys.argv[1], a, 48000, 'ogg'))\n"
"t.start(); t.join()\n"
)
r = subprocess.run([sys.executable, "-c", code, str(path)], capture_output=True, text=True)
assert r.returncode == 0, f"exit {r.returncode:#x}: {r.stderr[-400:]}"
audio, sr = post.read(path)
assert sr == 48000 and audio.shape[0] == 2
assert abs(audio.shape[-1] - 48000 * 60) < 2000
@needs_ffmpeg
def test_ffmpeg_resample():
out, info = post.ffmpeg_process(sine(1.0), 44100, 48000, None, -1.0)
assert out.shape[-1] == pytest.approx(48000, abs=10)
assert info == {}
@needs_ffmpeg
def test_ffmpeg_loudness():
out, info = post.ffmpeg_process(sine(5.0, amp=0.05), 44100, 48000, -16.0, -1.0)
assert info["output_lufs"] == pytest.approx(-16.0, abs=1.0)
def test_audible_end():
sr = 1000
a = np.concatenate([sine(2.0, sr=sr, freq=50), np.zeros((2, sr), np.float32)], axis=1)
assert post.audible_end(a, sr) == pytest.approx(2.0, abs=0.01)
assert post.audible_end(np.zeros((2, sr), np.float32), sr) == 0.0