"""
Transcribe course book listening-exercise audio to German text using
faster-whisper (GPU-accelerated via CTranslate2).

Setup (one time):
    pip install faster-whisper

Usage:
    python transcribe.py <path-to-audio-folder> <path-to-output-folder>

Example:
    python transcribe.py "D:/MenschenA1.1/audio/Lektion01" "../A1.1/course-book/transcripts/lektion-01"

Each audio file gets a matching .txt file with the same basename.
Review the output — ASR on textbook dialogue audio is generally good but can
mangle names, numbers, and fast/overlapping speech. Correct those before
committing.
"""

import sys
from pathlib import Path

from faster_whisper import WhisperModel

# "large-v3" is most accurate; drop to "medium" if VRAM is limited (~5GB vs ~10GB).
MODEL_SIZE = "large-v3"
DEVICE = "cuda"
COMPUTE_TYPE = "float16"  # use "int8_float16" if you run out of VRAM

AUDIO_EXTENSIONS = {".mp3", ".wav", ".m4a", ".ogg", ".flac"}


def main() -> None:
    if len(sys.argv) != 3:
        print(__doc__)
        sys.exit(1)

    audio_dir = Path(sys.argv[1])
    out_dir = Path(sys.argv[2])
    out_dir.mkdir(parents=True, exist_ok=True)

    model = WhisperModel(MODEL_SIZE, device=DEVICE, compute_type=COMPUTE_TYPE)

    audio_files = sorted(
        p for p in audio_dir.iterdir() if p.suffix.lower() in AUDIO_EXTENSIONS
    )
    if not audio_files:
        print(f"No audio files found in {audio_dir}")
        return

    for audio_path in audio_files:
        print(f"Transcribing {audio_path.name} ...")
        segments, info = model.transcribe(str(audio_path), language="de", beam_size=5)

        out_path = out_dir / (audio_path.stem + ".txt")
        with out_path.open("w", encoding="utf-8") as f:
            for segment in segments:
                f.write(f"[{segment.start:6.1f}s] {segment.text.strip()}\n")

        print(f"  -> {out_path}")

    print("\nDone. Review transcripts for ASR errors before committing.")


if __name__ == "__main__":
    main()