Voice cloning

# labs/glados_voice.py
import wave
from pathlib import Path

REF_FILE = "glados/data/voice/reference.wav"
REF_TEXT = "This is my reference recording for the local assistant."

def reference_info(ref_file: str) -> dict[str, float]:
    path = Path(ref_file)
    if not path.exists():
        raise FileNotFoundError(f"Reference clip not found: {ref_file}")
    with wave.open(str(path), "rb") as wav:
        return {"seconds": round(wav.getnframes() / wav.getframerate(), 2),
                "sample_rate": float(wav.getframerate())}

if __name__ == "__main__":
    info = reference_info(REF_FILE)
    print(f"Reference OK: {info['seconds']}s @ {int(info['sample_rate'])} Hz")

$ uv run python labs/glados_voice.py
Reference OK: 4.9s @ 24000 Hz

Use the recording you made or obtained with consent in chapter 6, and change REF_TEXT if you recorded different words. The retained run-outs used an earlier reference; their durations are exemplars, not measurements of your new file. Why check before loading? Because instantiating the cloning model downloads roughly two gigabytes and pins your machine for tens of seconds, and if the path is wrong or the file is not really a WAV you want to fail in milliseconds with a clear message instead. The standard-library wave module reads just the header: chapter 3 knowledge, now doing gatekeeping duty. Pick your reference deliberately (a clean four-to-ten-second clip, one consenting voice alone, no background music) and make sure REF_TEXT matches the audio word for word, punctuation included. The model aligns the reference audio against that text to learn the voice, and a sloppy transcript smears everything downstream.

from f5_tts.api import F5TTS

OUTPUT = "glados_says.wav"

def say(tts: F5TTS, text: str) -> float:
    wav, sr, _ = tts.infer(
        ref_file=REF_FILE,
        ref_text=REF_TEXT,
        gen_text=text,
        file_wave=OUTPUT,
        remove_silence=True,
    )
    return len(wav) / sr

Two details carry weight. The import is f5_tts.api, because the public interface lives in that submodule; import from the bare package and there is no F5TTS to be found (the failure box reproduces it). And infer() returns a three-tuple of waveform, sample rate and spectrogram; we keep two, discard the spectrogram with _, and return the duration. Wrapping the call in say() means nothing else in the program ever has to remember that signature — chapter 10 imports this exact function.

import sounddevice as sd
import soundfile as sf

def main() -> None:
    info = reference_info(REF_FILE)
    print(f"Reference OK: {info['seconds']}s @ {int(info['sample_rate'])} Hz")
    tts = F5TTS()
    print("Type something for GLaDOS to say (blank line to quit).")
    while True:
        text = input("GLaDOS > ").strip()
        if not text:
            break
        seconds = say(tts, text)
        print(f"Generated {seconds:.2f}s of audio. Playing...")
        data, samplerate = sf.read(OUTPUT)
        sd.play(data, samplerate)
        sd.wait()

if __name__ == "__main__":
    main()

$ uv run python labs/glados_voice.py
Reference OK: 4.9s @ 24000 Hz
Type something for GLaDOS to say (blank line to quit).
GLaDOS > You found this chapter. I'm genuinely impressed. And a little disappointed.
Generated 5.84s of audio. Playing...
GLaDOS >

The program checks the recording before loading the model, then reuses the loaded instance for each sentence. Measure generation time on your machine.