fix: correct G2P duration estimation and remove punctuation spaces

When G2P is enabled, estimate audio duration from the original Chinese
text instead of the pinyin output to avoid weight inflation from tone
number digits (weight 3.5 vs CJK 3.0). Also strip spaces around
punctuation in the G2P output.

app.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import logging
 import os
+import re
 from dataclasses import dataclass
 from typing import Any
 
@@ -158,7 +159,10 @@ def apply_g2p(text: str, dialect: str) -> str:
 
     lang_group = DIALECT_TO_LANG_GROUP.get(dialect, "hak_sx")
     result = g2p(text, lang_group=lang_group, pronunciation_type="pinyin")
-    return " ".join(result.pronunciations).upper()
+    joined = " ".join(result.pronunciations).upper()
+    joined = re.sub(r"\s+([，。！？；：、…「」『』【】〔〕（）])", r"\1", joined)
+    joined = re.sub(r"([，。！？；：、…「」『』【】〔〕（）])\s+", r"\1", joined)
+    return joined
 
 
 def validate_inputs(
@@ -208,11 +212,9 @@ def synthesize(
         return None, startup_status()
 
     try:
-        input_text = text.strip()
+        original_text = text.strip()
         g2p_note = ""
-        if use_g2p:
-            input_text = apply_g2p(input_text, dialect)
-            g2p_note = f"；G2P 轉換：{input_text}"
+        duration_override = None
 
         generation_config = RUNTIME.generation_config_cls(
             num_step=int(num_step),
@@ -226,6 +228,21 @@ def synthesize(
             ref_text=ref_text.strip(),
             preprocess_prompt=True,
         )
+
+        if use_g2p:
+            input_text = apply_g2p(original_text, dialect)
+            g2p_note = f"；G2P 轉換：{input_text}"
+            # Estimate duration from original Chinese text to avoid weight inflation
+            # caused by tone number digits (weight 3.5) in the G2P output.
+            num_ref_tokens = voice_clone_prompt.ref_audio_tokens.size(-1)
+            frame_rate = RUNTIME.model.audio_tokenizer.config.frame_rate
+            est_frames = RUNTIME.model.duration_estimator.estimate_duration(
+                original_text, voice_clone_prompt.ref_text, num_ref_tokens
+            )
+            duration_override = est_frames / float(speed) / frame_rate
+        else:
+            input_text = original_text
+
         generate_kwargs: dict[str, Any] = {
             "text": input_text,
             "voice_clone_prompt": voice_clone_prompt,
@@ -233,7 +250,9 @@ def synthesize(
             "generation_config": generation_config,
             "language": "zh",
         }
-        if speed != DEFAULT_SPEED:
+        if duration_override is not None:
+            generate_kwargs["duration"] = duration_override
+        elif speed != DEFAULT_SPEED:
             generate_kwargs["speed"] = float(speed)
 
         audio = RUNTIME.model.generate(**generate_kwargs)