Update TTS server backend and model weights via upload script

app.py

@@ -20,7 +20,7 @@ BASE_DIR = Path(__file__).resolve().parent
 DEFAULT_MODEL = BASE_DIR / "models" / "lumua.onnx"
 SAMPLE_RATE = 22050
 VOCAB_SIZE = 10000
-MAX_CHARS = 200  # per chunk, keeps each call well under the ~510 token limit
+MAX_CHARS = 600  # per chunk, keeps each call well under the ~510 token limit
 
 BOS_ID = 1
 EOS_ID = 2
@@ -90,7 +90,7 @@ def get_session(model_path: str) -> ort.InferenceSession:
 
 
 def split_text(text: str, max_chars: int = MAX_CHARS) -> list[str]:
-    sentences = re.split(r"(?<=[.!?;:])\s+", text.strip())
+    sentences = re.split(r"(?<=[.!?])\s+", text.strip())
     chunks = []
     for sentence in sentences:
         sentence = sentence.strip()
@@ -111,6 +111,20 @@ def split_text(text: str, max_chars: int = MAX_CHARS) -> list[str]:
     return chunks
 
 
+FADE = int(SAMPLE_RATE * 0.01)
+
+
+def fade_edges(x: np.ndarray) -> np.ndarray:
+    n = min(FADE, len(x) // 2)
+    if n == 0:
+        return x
+    x = x.copy()
+    ramp = np.linspace(0.0, 1.0, n, dtype=np.float32)
+    x[:n] *= ramp
+    x[-n:] *= ramp[::-1]
+    return x
+
+
 def synthesize(text, speed, pause, model_path):
     if not text or not text.strip():
         raise gr.Error("Please enter some text.")
@@ -129,7 +143,7 @@ def synthesize(text, speed, pause, model_path):
             continue
         tokens = np.array([ids], dtype=np.int64)
         wav = session.run(None, {"tokens": tokens, "speed": speed_arr})[0]
-        pieces.append(np.asarray(wav, dtype=np.float32).squeeze())
+        pieces.append(fade_edges(np.asarray(wav, dtype=np.float32).squeeze()))
         pieces.append(gap)
     elapsed = time.perf_counter() - start