Update app.py

app.py

@@ -5,6 +5,7 @@ import soundfile as sf
 from xcodec2.modeling_xcodec2 import XCodec2Model
 import torchaudio
 import gradio as gr
+import re
 
 llasa_model_id = 'OmniAICreator/Llasa-1B-run1'
 
@@ -28,6 +29,68 @@ whisper_turbo_pipe = pipeline(
     device='cuda',
 )
 
+REPLACE_MAP: dict[str, str] = {
+    r"\t": "",
+    r"\[n\]": "",
+    r" ": "",
+    r"　": "",
+    r"[;▼♀♂《》≪≫①②③④⑤⑥]": "",
+    r"[\u02d7\u2010-\u2015\u2043\u2212\u23af\u23e4\u2500\u2501\u2e3a\u2e3b]": "",
+    r"[\uff5e\u301C]": "ー",
+    r"？": "?",
+    r"！": "!",
+    r"[●◯〇]": "○",
+    r"♥": "♡",
+}
+FULLWIDTH_ALPHA_TO_HALFWIDTH = str.maketrans(
+    {
+        chr(full): chr(half)
+        for full, half in zip(
+            list(range(0xFF21, 0xFF3B)) + list(range(0xFF41, 0xFF5B)),
+            list(range(0x41, 0x5B)) + list(range(0x61, 0x7B)),
+        )
+    }
+)
+HALFWIDTH_KATAKANA_TO_FULLWIDTH = str.maketrans(
+    {
+        chr(half): chr(full)
+        for half, full in zip(range(0xFF61, 0xFF9F), range(0x30A1, 0x30FB))
+    }
+)
+FULLWIDTH_DIGITS_TO_HALFWIDTH = str.maketrans(
+    {
+        chr(full): chr(half)
+        for full, half in zip(range(0xFF10, 0xFF1A), range(0x30, 0x3A))
+    }
+)
+INVALID_PATTERN = re.compile(
+    r"[^\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF\u3400-\u4DBF\u3005"
+    r"\u0041-\u005A\u0061-\u007A"
+    r"\u0030-\u0039"
+    r"。、!?…♪♡○]"
+)
+
+def normalize(text: str) -> str:
+    for pattern, replacement in REPLACE_MAP.items():
+        text = re.sub(pattern, replacement, text)
+
+    text = text.translate(FULLWIDTH_ALPHA_TO_HALFWIDTH)
+    text = text.translate(FULLWIDTH_DIGITS_TO_HALFWIDTH)
+    text = text.translate(HALFWIDTH_KATAKANA_TO_FULLWIDTH)
+
+    text = re.sub(r"…{2,}", "…", text)
+    text = re.sub(r"ー{2,}", "ー", text)
+
+    def replace_special_chars(match):
+        seq = match.group(0)
+        return seq[0] if len(set(seq)) == 1 else seq[0] + seq[-1]
+
+    text = re.sub(r"[!?♪♡]{2,}", replace_special_chars, text)
+
+    repeated = "ッっあいうえおんぁぃぅぇぉゃゅょアイウエオンァィゥェォャュョ"
+    text = re.sub(f"([{repeated}])\\1{{2,}}", r"\1\1", text)
+    return text
+
 def ids_to_speech_tokens(speech_ids):
  
     speech_tokens_str = []
@@ -56,6 +119,7 @@ def infer(sample_audio_path, target_text, temperature, top_p,  progress=gr.Progr
     if len(target_text) > 300:
         gr.Warning("Text is too long. Please keep it under 300 characters.")
         target_text = target_text[:300]
+    target_text = normalize(target_text)
     with torch.no_grad():
         if sample_audio_path:
             progress(0, 'Loading and trimming audio...')