preprocess_text

vits/vits_inferencer.py

@@ -24,13 +24,32 @@ default_noise_scale = 0.667
 default_noise_scale_w = 0.8
 default_length_scale = 1
 
+replace_list = [
+    ("candle", "刊豆"),
+    ("end", "按的"),
+    ("hoshimi", "吼西咪"),
+    ("mua", "木啊"),
+    ("hsm", "吼西咪"),
+    ("ho", "齁"),
+    ("na", "呐"),
+    ("shi", "西"),
+    ("mi", "咪"),
+]
+
 def get_text(text, hps):
+    text = preprocess_text(text)
     text_norm = text_to_sequence(text, hps.data.text_cleaners)
     if hps.data.add_blank:
         text_norm = commons.intersperse(text_norm, 0)
     text_norm = torch.LongTensor(text_norm)
     return text_norm
 
+def preprocess_text(text):
+    text = text.lower()
+    for src, dst in replace_list:
+        text = text.replace(src, dst)
+    return text
+
 class VitsInferencer:
     def __init__(self, hps_path, device="cpu"):
         print("init")
@@ -65,8 +84,8 @@ class VitsInferencer:
         length_scale = gr.Slider(minimum=0, maximum=3, value=default_length_scale, step=0.001, label="length_scale（数值越大输出音频越长）")
         
         tts_input = gr.TextArea(
-            label="请输入文本（目前只支持汉字和单个英文字母，可以使用常用符号和空格来改变语调和停顿，请勿一次性输入过长文本）",
-            value="这里是爱喝奶茶，穿得也像奶茶魅力点是普通话二乙的星弥吼西咪，晚上齁。")
+            label="请输入文本（目前只支持汉字、单个英文字母和极个别专有名词，可以使用常用符号和空格来改变语调和停顿，请勿一次性输入过长文本）",
+            value="这里是爱喝奶茶，穿得也像奶茶魅力点是普通话二乙的星弥Hoshimi，晚上Ho")
         tts_submit = gr.Button("合成", variant="primary")
         tts_output = gr.Audio(label="Output")
         gr.HTML('''