add mel sharpen (before soft-gate): restore aspiration detail

python/infer_board.py

@@ -129,6 +129,21 @@ def mel_soft_gate(mel, thr_lo=MEL_GATE_LO, thr_hi=MEL_GATE_HI, floor=-11.5):
     return out
 
 
+def mel_sharpen(mel, alpha=0.5, k=5):
+    """Spectral contrast boost along the mel-band axis: m + alpha*(m - smooth(m))."""
+    if alpha <= 0 or k <= 1:
+        return mel
+    m = mel[0] if mel.ndim == 3 else mel
+    pad = k // 2
+    mp = np.pad(m, ((pad, pad), (0, 0)), mode="reflect")
+    smooth = np.zeros_like(m)
+    for i in range(k):
+        smooth += mp[i:i + m.shape[0]]
+    smooth /= k
+    sharp = m + alpha * (m - smooth)
+    return (sharp if mel.ndim == 2 else sharp[None]).astype(np.float32)
+
+
 def tail_stretch(mel, n_old=TAIL_STRETCH_OLD, n_new=TAIL_STRETCH_NEW):
     """句尾 mel 拉伸：最后 n_old 帧线性插值到 n_new 帧。
     模型对句尾音节 duration 预测偏短（如 11 帧 vs 参考 43 帧），
@@ -212,6 +227,7 @@ def synthesize(acoustic, vocoder, text, noise_scale=0.3, seed=0):
     for s in sents:
         x, x_lengths, noise_z = text_to_inputs(s, noise_scale, seed)
         mel = acoustic_to_mel(acoustic, x, x_lengths, noise_z)
+        mel = mel_sharpen(mel)
         mel = mel_soft_gate(mel)
         mel = tail_stretch(mel)
         pieces.append(vocoder_chunked(vocoder, mel))