"""Old Tam narrates the trailer (tools/trailer/make_trailer.sh mixes it in):
one line for each shot, in his voice (the same clone as
tools/voice/make_voice.py), each checked by Whisper, the best of three takes
kept, and sped up a touch if it would run past the start of the next line.
Output: tools/trailer/voice/<NN>.ogg and lines.json ([{file, shot, at, text}]:
`at` is seconds into its shot; make_trailer.sh reads the shots' start times
from the render log). Run with the voice environment:
  tools/voice/.venv-qwen/bin/python tools/trailer/make_trailer_voice.py
"""
import json, os, subprocess, sys, tempfile

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.join(HERE, "..", "voice"))
import make_voice as mv   # (its voice, its checks)

OUT = os.path.join(HERE, "voice")
## The shots (trailer.gd), their lengths in seconds, and what he says over
## each, starting `at` seconds in.
SHOTS = [7.5, 5.0, 6.0, 5.5, 4.5, 5.5, 6.0, 6.0]
LINES = [
    (0, 2.0, "Welcome to Cinderwild. You wake on a wild island, washed up with nothing but the clothes on your back."),
    (1, 0.3, "Gather what you can. Hunt, forage, and cook over an open fire."),
    (2, 0.3, "Build yourself a home, plank by plank, before the nights draw in."),
    (3, 0.3, "Because when the sun goes down, the wolves come out. Keep your fire burning."),
    (4, 0.3, "Weather the storms, and the hard winters."),
    (5, 0.3, "Then dig down, through the sea cliffs."),
    (6, 0.3, "And find out what lies beyond."),
    (7, 0.6, "Play alone, or with your friends, on your own server. Cinderwild."),
]
GAP = 0.25   # seconds of quiet before the next line


def main():
    import numpy as np
    import soundfile as sf
    import mlx_whisper
    from mlx_audio.tts.utils import load_model
    os.makedirs(OUT, exist_ok=True)
    for f in os.listdir(OUT):
        if f.endswith(".ogg"):
            os.remove(os.path.join(OUT, f))
    model = load_model(mv.MODEL)
    sr = getattr(model, "sample_rate", 24000)
    tmp = tempfile.mkdtemp()
    out = []
    for n, (shot, at, text) in enumerate(LINES):
        # Room: to the next line's start (or the end of the film).
        if n + 1 < len(LINES):
            nshot, nat, _ = LINES[n + 1]
            room = sum(SHOTS[shot:nshot]) - at + nat - GAP
        else:
            room = sum(SHOTS[shot:]) - at - 0.5
        best = None
        for k in range(3):
            res = list(model.generate(text=text, ref_audio=mv.REF_AUDIO, ref_text=mv.REF_TEXT))
            audio = np.concatenate([np.array(r.audio) for r in res])
            wav = os.path.join(tmp, "%02d_%d.wav" % (n, k))
            sf.write(wav, audio, sr)
            heard = mlx_whisper.transcribe(wav, path_or_hf_repo=mv.WHISPER, language="en")["text"].strip()
            secs = len(audio) / sr
            # Words first; then the one that fits.
            score = mv.match(text, heard) - (0.05 if secs > room * 1.15 else 0.0)
            if best is None or score > best[0]:
                best = (score, wav, secs)
            if score >= 0.98 and secs <= room:
                break
        score, wav, secs = best
        tempo = min(max(secs / room, 1.0), 1.2) if room > 0 else 1.0
        name = "%02d.ogg" % n
        subprocess.run(["ffmpeg", "-v", "error", "-y", "-i", wav, "-af", "atempo=%.3f" % tempo, "-ac", "2",
                        "-c:a", "vorbis", "-strict", "-2", "-q:a", "5", os.path.join(OUT, name)], check=True)
        out.append({"file": name, "shot": shot, "at": at, "text": text})
        print("  %s shot %d: %.1f s%s (room %.1f s, words %.0f%%)" % (name, shot, secs / tempo,
              " (x%.2f)" % tempo if tempo > 1.0 else "", room, score * 100), flush=True)
    json.dump(out, open(os.path.join(OUT, "lines.json"), "w"), indent=2)


if __name__ == "__main__":
    main()