tests/fixtures/make_fixtures.py

"""Generate the non-speech audio fixtures. Deterministic: fixed seed, fixed params.

Chromium's --use-file-for-fake-audio-capture requires 16-bit PCM WAV.
VOICEVOX outputs 24000 Hz mono, so match it for consistency across fixtures.
"""

import wave
from pathlib import Path

import numpy as np

SR = 24000
SECONDS = 30
HERE = Path(__file__).parent


def _write(path: Path, samples: np.ndarray) -> None:
    pcm = np.clip(samples, -1.0, 1.0)
    pcm = (pcm * 32767).astype("<i2")
    with wave.open(str(path), "wb") as w:
        w.setnchannels(1)
        w.setsampwidth(2)
        w.setframerate(SR)
        w.writeframes(pcm.tobytes())


def make_silence() -> Path:
    p = HERE / "silence_30s.wav"
    _write(p, np.zeros(SR * SECONDS, dtype=np.float64))
    return p


def make_cafe_noise() -> Path:
    """Pink-ish noise at conversational level: broadband, speech-shaped, no speech.

    Deliberately loud (measured RMS 0.0577, i.e. -24.8 dBFS; peak 0.1357). It must be
    well ABOVE any plausible RMS floor, so
    that plan 01-07's gate is forced to discriminate speech from noise on envelope
    modulation rather than on loudness. A quiet noise fixture would let a naive
    RMS-only gate pass the test for the wrong reason.
    """
    rng = np.random.default_rng(20260826)
    white = rng.standard_normal(SR * SECONDS)
    pink = np.cumsum(white)
    pink = pink - np.mean(pink)
    pink = pink / (np.max(np.abs(pink)) or 1.0)
    mixed = 0.7 * pink + 0.3 * (white / (np.max(np.abs(white)) or 1.0))
    p = HERE / "cafe_noise_30s.wav"
    _write(p, mixed * 0.15)
    return p


def wav_duration_seconds(path: Path) -> float:
    with wave.open(str(path), "rb") as w:
        return w.getnframes() / float(w.getframerate())


if __name__ == "__main__":
    for fn in (make_silence, make_cafe_noise):
        out = fn()
        print(f"{out.name}: {wav_duration_seconds(out):.3f}s")