Her Own Name

# labs/wake_word.py
WAKE_WORDS = ["glados", "hey glados", "aperture"]

def contains_wake_word(text: str) -> bool:
    text = text.lower()
    return any(word in text for word in WAKE_WORDS)

print(contains_wake_word("Hey GLaDOS, are you there?"))
print(contains_wake_word("what time is it"))
print(contains_wake_word("Welcome to Aperture Science"))

$ uv run python labs/wake_word.py
True
False
True

Substring matching, not equality, because Whisper returns sentences: "hey glados what's up," never a bare keyword. Lowercasing first absorbs Whisper's unpredictable capitalization of sentence starts and proper nouns. And the phrase list is configurable data, not constants scattered through logic; you have read enough of this book to have expected that.

import numpy as np

SILENCE_THRESHOLD = 0.005

def rms_energy(audio: np.ndarray) -> float:
    return float(np.sqrt(np.mean(audio ** 2)))

def is_silence(audio: np.ndarray) -> bool:
    return rms_energy(audio) < SILENCE_THRESHOLD

quiet = np.full(32000, 0.001, dtype="float32")
loud = np.random.default_rng(0).uniform(-0.3, 0.3, 32000).astype("float32")
print(f"quiet rms={rms_energy(quiet):.4f}  silence={is_silence(quiet)}")
print(f"loud  rms={rms_energy(loud):.4f}  silence={is_silence(loud)}")

$ uv run python labs/wake_word.py
quiet rms=0.0010  silence=True
loud  rms=0.1726  silence=False

Chapter 4's RMS, promoted from a debugging guard to a load-bearing filter. The threshold sits at 0.005 rather than zero because a real microphone never reads exactly zero; there is always a noise floor, and the threshold lives above it and below speech. The synthetic buffers make the test deterministic (a constant near-silent hum and a seeded burst of noise) so this stage runs identically on any machine, no microphone required. Your live threshold may want tuning against the levels you measured in chapter 4's exercise 2; that number was for exactly this.

import sounddevice as sd
import soundfile as sf
from faster_whisper import WhisperModel

SAMPLE_RATE = 16000
CHANNELS = 1
CHUNK_DURATION = 2
AUDIO_FILE = "glados/data/wake_chunk.wav"

def detect_wake_word(model: WhisperModel) -> tuple[bool, str]:
    audio = sd.rec(
        int(CHUNK_DURATION * SAMPLE_RATE),
        samplerate=SAMPLE_RATE,
        channels=CHANNELS,
        dtype="float32",
    )
    sd.wait()
    if is_silence(audio):
        return False, ""
    sf.write(AUDIO_FILE, audio, SAMPLE_RATE)
    segments, _ = model.transcribe(AUDIO_FILE, language="en")
    text = " ".join(s.text.strip().lower() for s in segments)
    return contains_wake_word(text), text

if __name__ == "__main__":
    model = WhisperModel("base", device="cpu", compute_type="int8")
    print("Waiting for wake word. Say 'Hey GLaDOS'...")
    while True:
        woke, heard = detect_wake_word(model)
        if woke:
            print(f"Wake word detected in: {heard!r}")
            break
        if heard:
            print(f"(heard, ignored: {heard!r})")

$ uv run python labs/wake_word.py
Waiting for wake word. Say 'Hey GLaDOS'...
(heard, ignored: 'is this thing even on')
Wake word detected in: 'hey glados are you awake'

Read the function's ordering like a bill: record two seconds, then the free check, and only if energy is present do we pay for the disk write and the transcription. The return type carries both the verdict and the transcript, so a caller can log what was ignored — which is how you discover your threshold is eating quiet speech, or that the cat's collar bell reads as audio. The loop prints the ignored lines for exactly that reason: what a filter rejected tells you more than what it accepted.