| """VoxCPM2 TTS — uses the exact API from openbmb/VoxCPM2 model card.
|
| Voice design is achieved by prefixing text with '(voice description) actual text'.
|
| """
|
| import os, uuid, numpy as np, soundfile as sf
|
| from pathlib import Path
|
| from ..config import AUDIO_CACHE
|
|
|
| _ENG = None
|
|
|
|
|
| def _voice_clause(vd: dict) -> str:
|
| pace = vd.get("pace", "moderate")
|
| tone = vd.get("tone", "calm_grounded").replace("_", " ")
|
| vid = vd.get("voice_id", "")
|
|
|
| gender = "female" if "guardian" in vid or "weaver" in vid or "presenter" in vid else \
|
| "male" if "builder" in vid or "fetcher" in vid else "neutral"
|
| return f"({gender} voice, {tone} tone, {pace} pace)"
|
|
|
|
|
| def _lazy():
|
| global _ENG
|
| if _ENG is not None:
|
| return _ENG
|
| try:
|
| from voxcpm import VoxCPM
|
| _ENG = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)
|
| print("[voxcpm] loaded")
|
| except Exception as e:
|
| print(f"[voxcpm] load failed ({e}) — using silent fallback")
|
| _ENG = "SILENT"
|
| return _ENG
|
|
|
|
|
| def synthesize(text: str, voice_design: dict) -> str:
|
| eng = _lazy()
|
| out = AUDIO_CACHE / f"{uuid.uuid4().hex}.wav"
|
| if not text.strip():
|
| sf.write(out, np.zeros(int(48000 * 0.3), dtype=np.float32), 48000)
|
| return str(out)
|
| if eng == "SILENT" or eng is None:
|
| sf.write(out, np.zeros(int(48000 * 0.6), dtype=np.float32), 48000)
|
| return str(out)
|
| prefix = _voice_clause(voice_design or {})
|
| full_text = f"{prefix}{text}"
|
| try:
|
| wav = eng.generate(text=full_text, cfg_value=2.0, inference_timesteps=10)
|
| sr = eng.tts_model.sample_rate
|
| sf.write(out, wav, sr)
|
| except Exception as e:
|
| print(f"[voxcpm] generation failed: {e}")
|
| sf.write(out, np.zeros(int(48000 * 0.6), dtype=np.float32), 48000)
|
| return str(out)
|
|
|
