Upload comfy/custom_nodes/ComfyUI-IndexTTS2/audition_klaus.py with huggingface_hub

comfy/custom_nodes/ComfyUI-IndexTTS2/audition_klaus.py

@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+"""audition_klaus.py — single-fragment IndexTTS-2 audition for the Web UI's
+歪果仁 mode试听 button.
+
+Loads IndexTTS-2 once, runs ONE infer with klaus_lexicon mangling on the
+provided text, writes a single FLAC. Sibling to render_klaus_episode.py but
+single-fragment + no concat / atempo / sidecar — the Web UI just needs a
+~6-second self-introduction sample to preview the voice.
+
+Cold-start ~4 min (model load); warm ~3-5 s. The Web UI cache-by-name keeps
+this cost paid only once per voice — see voice_workshop.audition_indextts.
+
+Usage (server-side):
+    cd /root/IndexTTS2
+    .venv/bin/python audition_klaus.py \
+        --ref-wav refs/web/<voice>.flac \
+        --text "嗨，你好呀！我是 Klaus..." \
+        --output /root/ComfyUI/output/voice_audition/klaus_audition_klaus_NNN.flac
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent))
+import klaus_lexicon  # noqa: E402
+
+from indextts.infer_v2 import IndexTTS2  # noqa: E402
+
+# Same locked-in profile as render_klaus_episode.py — keeps the audition
+# faithful to what the actual render will sound like.
+LEXICON_PROFILE = dict(
+    tone_4to1_prob=0.95,
+    tone_3to2_prob=0.95,
+    retroflex_loss_prob=0.0,
+    u_prob=0.0,
+)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
+    ap.add_argument("--ref-wav", required=True, help="path to spk_audio_prompt (relative to cwd or absolute)")
+    ap.add_argument("--text", required=True, help="self-intro text to speak")
+    ap.add_argument("--output", required=True, help="final .flac path")
+    ap.add_argument("--seed", type=int, default=42)
+    args = ap.parse_args()
+
+    if not Path(args.ref_wav).exists():
+        sys.exit(f"REF missing: {args.ref_wav} (cwd={Path.cwd()})")
+
+    print(">> loading IndexTTS2…", flush=True)
+    t0 = time.perf_counter()
+    tts = IndexTTS2(cfg_path="checkpoints/config.yaml", model_dir="checkpoints", use_fp16=True)
+    print(f">> loaded in {time.perf_counter()-t0:.1f}s on {tts.device}", flush=True)
+
+    mangled, stats = klaus_lexicon.transform(args.text, seed=args.seed, **LEXICON_PROFILE)
+    rate = stats["overrides_emitted"] / max(stats["total_han"], 1)
+    print(f">> mangled ({rate:.0%} override): {mangled}", flush=True)
+
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+    t0 = time.perf_counter()
+    tts.infer(spk_audio_prompt=args.ref_wav, text=mangled, output_path=args.output, verbose=False)
+    print(f">> done in {time.perf_counter()-t0:.1f}s -> {args.output}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())