Old Tam in the trailer: three short lines, timed to the dawn flyover, the wolves and the Cove (same clip size)

tools/trailer/make_trailer_voice.py

@@ -0,0 +1,51 @@
+"""Old Tam's lines for the trailer (tools/trailer/make_trailer.sh mixes them
+in): short ones, in his voice (the same clone as tools/voice/make_voice.py),
+each checked by Whisper and the best of three takes kept. Output:
+tools/trailer/voice/<name>.ogg. 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")
+## name -> what he says (make_trailer.sh times each against the shots).
+LINES = {
+    "dawn": "Ha! Awake at last! Washed up on my island, eh?",
+    "wolves": "Hush now. See it, there, at the edge of the light?",
+    "cove": "Well, you made it down!",
+}
+
+
+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)
+    model = load_model(mv.MODEL)
+    sr = getattr(model, "sample_rate", 24000)
+    tmp = tempfile.mkdtemp()
+    for name, text in LINES.items():
+        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, "%s_%d.wav" % (name, k))
+            sf.write(wav, audio, sr)
+            heard = mlx_whisper.transcribe(wav, path_or_hf_repo=mv.WHISPER, language="en")["text"].strip()
+            score = mv.match(text, heard)
+            if best is None or score > best[0]:
+                best = (score, wav, len(audio) / sr)
+            if score >= 0.98:
+                break
+        subprocess.run(["ffmpeg", "-v", "error", "-y", "-i", best[1], "-ac", "2", "-c:a", "vorbis", "-strict", "-2",
+                        "-q:a", "5", os.path.join(OUT, name + ".ogg")], check=True)
+        print("  %s (%.1f s, words %.0f%%)" % (name, best[2], best[0] * 100), flush=True)
+    json.dump(LINES, open(os.path.join(OUT, "lines.json"), "w"), indent=2)
+
+
+if __name__ == "__main__":
+    main()