DreamVoice: ZeroGPU app

tts.py

@@ -0,0 +1,125 @@
+"""VoxCPM2 voice cloning and narration — local GPU inference for HF Spaces."""
+
+from __future__ import annotations
+
+import os
+import re
+import tempfile
+
+import numpy as np
+
+MODEL_ID = "openbmb/VoxCPM2"
+
+_model = None
+
+
+def _get_model():
+    global _model
+    if _model is None:
+        from voxcpm import VoxCPM
+        _model = VoxCPM.from_pretrained(
+            MODEL_ID,
+            device="cuda",
+            load_denoiser=True,
+            optimize=True,
+        )
+    return _model
+
+
+# Load at module level for ZeroGPU (CUDA emulation outside @spaces.GPU)
+try:
+    _get_model()
+except Exception:
+    pass
+
+
+def _split_sentences(text: str, max_chars: int = 200):
+    parts = re.split(r"(?<=[.!?।])\s+|\n+", text.strip())
+    out = []
+    for p in parts:
+        p = p.strip()
+        if not p:
+            continue
+        while len(p) > max_chars:
+            cut = p.rfind(" ", 0, max_chars)
+            cut = cut if cut > 0 else max_chars
+            out.append(p[:cut].strip())
+            p = p[cut:].strip()
+        out.append(p)
+    return out or [text.strip()]
+
+
+def _pause_for(mood: str, energy: float = 0.45) -> float:
+    energy = max(0.0, min(1.0, float(energy)))
+    base = 0.45 if mood in ("funny", "magical") else 0.65
+    return round(base + (0.85 - base) * (1.0 - energy), 3)
+
+
+def _with_bedtime_style(text: str, speed: float, mood: str = "", energy: float = 0.45) -> str:
+    energy = max(0.0, min(1.0, float(energy)))
+    mood_styles = {
+        "magical": "gentle, warm, slightly slow, wonder-filled whisper, like telling a secret about something beautiful, soft rising intonation on wonder words, pause briefly after each beat",
+        "funny": "warm, playful, slightly animated, gentle humor in the voice, light chuckle between lines, bright and cheerful but still soft enough for bedtime, slightly faster pace than usual",
+        "calming": "very slow, deep warm whisper, barely above a breath, each word drifting gently into the next, long pauses between sentences, voice fading softly at the end of each line, like someone falling asleep while reading",
+        "dreamy": "slow, soft, breathy whisper, voice drifting like floating on a cloud, elongated vowels, gentle hum between phrases, lullaby-like rhythm, words dissolving into silence",
+    }
+    style = mood_styles.get(mood, "gentle, warm, sleepy bedtime voice, slightly slow pace")
+    if energy >= 0.66:
+        style += ", a little brighter and more animated, lively for a delighted child"
+    elif energy <= 0.33:
+        style += ", even softer and slower, barely above a whisper"
+    return f"({style}){text}"
+
+
+def _postprocess_np(audio, sr):
+    from audio_postprocess import postprocess
+    return postprocess(audio, sr)
+
+
+def clone_and_speak(ref_wav: str, text: str, speed: float = 0.9, mood: str = "", energy: float = 0.45) -> str:
+    """Clone the reference voice and synthesize text to a temporary WAV path.
+
+    Runs on local GPU (HF Spaces GPU Zero).
+    """
+    if not ref_wav or not os.path.exists(ref_wav):
+        raise ValueError("Please provide a prepared voice reference WAV.")
+
+    story_text = (text or "").strip()
+    if not story_text:
+        raise ValueError("Please provide story text to narrate.")
+
+    model = _get_model()
+    sr = int(model.tts_model.sample_rate)
+
+    pause = _pause_for(mood, energy)
+    silence = np.zeros(int(pause * sr), dtype=np.float32)
+
+    chunks = []
+    for sentence in _split_sentences(story_text):
+        wav = model.generate(
+            text=_with_bedtime_style(sentence, speed, mood, energy),
+            reference_wav_path=ref_wav,
+            cfg_value=2.0,
+            inference_timesteps=10,
+            normalize=True,
+            denoise=True,
+            retry_badcase=True,
+            retry_badcase_max_times=3,
+            retry_badcase_ratio_threshold=8.0,
+        )
+        wav = np.asarray(wav, dtype=np.float32)
+        if wav.size:
+            chunks.append(wav)
+            chunks.append(silence)
+
+    if not chunks:
+        raise RuntimeError("VoxCPM2 produced no audio.")
+
+    full = np.concatenate(chunks)
+    full = _postprocess_np(full, sr)
+
+    import soundfile as sf
+    fd, out_path = tempfile.mkstemp(prefix="dreamvoice_story_", suffix=".wav")
+    os.close(fd)
+    sf.write(out_path, full, sr)
+    return out_path