Merge https://github.com/RVC-Boss/GPT-SoVITS

GPT_SoVITS/inference_webui.py

@@ -4,12 +4,30 @@ logging.getLogger("urllib3").setLevel(logging.ERROR)
 logging.getLogger("httpcore").setLevel(logging.ERROR)
 logging.getLogger("httpx").setLevel(logging.ERROR)
 logging.getLogger("asyncio").setLevel(logging.ERROR)
+
+logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
+logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
 import pdb
-os.environ["no_proxy"] = "localhost, 127.0.0.1, ::1"
+
+if os.path.exists("./gweight.txt"):
+    with open("./gweight.txt", 'r',encoding="utf-8") as file:
+        gweight_data = file.read()
         gpt_path = os.environ.get(
-    "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
+    "gpt_path", gweight_data)
-)
+else:
+    gpt_path = os.environ.get(
+    "gpt_path", "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt")
+
+if os.path.exists("./sweight.txt"):
+    with open("./sweight.txt", 'r',encoding="utf-8") as file:
+        sweight_data = file.read()
+        sovits_path = os.environ.get("sovits_path", sweight_data)
+else:
     sovits_path = os.environ.get("sovits_path", "GPT_SoVITS/pretrained_models/s2G488k.pth")
+# gpt_path = os.environ.get(
+#     "gpt_path", "pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
+# )
+# sovits_path = os.environ.get("sovits_path", "pretrained_models/s2G488k.pth")
 cnhubert_base_path = os.environ.get(
     "cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base"
 )
@@ -62,7 +80,7 @@ def get_bert_feature(text, word2ph):
     with torch.no_grad():
         inputs = tokenizer(text, return_tensors="pt")
         for i in inputs:
-            inputs[i] = inputs[i].to(device)  #####输入是long不用管精度问题，精度随bert_model
+            inputs[i] = inputs[i].to(device)
         res = bert_model(**inputs, output_hidden_states=True)
         res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
     assert len(word2ph) == len(text)
@@ -119,6 +137,7 @@ def change_sovits_weights(sovits_path):
         n_speakers=hps.data.n_speakers,
         **hps.model
     )
+    if("pretrained"not in sovits_path):
         del vq_model.enc_q
     if is_half == True:
         vq_model = vq_model.half().to(device)
@@ -126,6 +145,7 @@ def change_sovits_weights(sovits_path):
         vq_model = vq_model.to(device)
     vq_model.eval()
     print(vq_model.load_state_dict(dict_s2["weight"], strict=False))
+    with open("./sweight.txt","w",encoding="utf-8")as f:f.write(sovits_path)
 change_sovits_weights(sovits_path)
 
 def change_gpt_weights(gpt_path):
@@ -142,6 +162,7 @@ def change_gpt_weights(gpt_path):
     t2s_model.eval()
     total = sum([param.nelement() for param in t2s_model.parameters()])
     print("Number of parameter: %.2fM" % (total / 1e6))
+    with open("./gweight.txt","w",encoding="utf-8")as f:f.write(gpt_path)
 change_gpt_weights(gpt_path)
 
 def get_spepc(hps, filename):
@@ -167,7 +188,84 @@ dict_language={
 }
 
 
-def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language):
+def splite_en_inf(sentence, language):
+    pattern = re.compile(r'[a-zA-Z. ]+')
+    textlist = []
+    langlist = []
+    pos = 0
+    for match in pattern.finditer(sentence):
+        start, end = match.span()
+        if start > pos:
+            textlist.append(sentence[pos:start])
+            langlist.append(language)
+        textlist.append(sentence[start:end])
+        langlist.append("en")
+        pos = end
+    if pos < len(sentence):
+        textlist.append(sentence[pos:])
+        langlist.append(language)
+
+    return textlist, langlist
+
+
+def clean_text_inf(text, language):
+    phones, word2ph, norm_text = clean_text(text, language)
+    phones = cleaned_text_to_sequence(phones)
+
+    return phones, word2ph, norm_text
+
+
+def get_bert_inf(phones, word2ph, norm_text, language):
+    if language == "zh":
+        bert = get_bert_feature(norm_text, word2ph).to(device)
+    else:
+        bert = torch.zeros(
+            (1024, len(phones)),
+            dtype=torch.float16 if is_half == True else torch.float32,
+        ).to(device)
+
+    return bert
+
+
+def nonen_clean_text_inf(text, language):
+    textlist, langlist = splite_en_inf(text, language)
+    phones_list = []
+    word2ph_list = []
+    norm_text_list = []
+    for i in range(len(textlist)):
+        lang = langlist[i]
+        phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
+        phones_list.append(phones)
+        if lang == "en" or "ja":
+            pass
+        else:
+            word2ph_list.append(word2ph)
+        norm_text_list.append(norm_text)
+    print(word2ph_list)
+    phones = sum(phones_list, [])
+    word2ph = sum(word2ph_list, [])
+    norm_text = ' '.join(norm_text_list)
+
+    return phones, word2ph, norm_text
+
+
+def nonen_get_bert_inf(text, language):
+    textlist, langlist = splite_en_inf(text, language)
+    print(textlist)
+    print(langlist)
+    bert_list = []
+    for i in range(len(textlist)):
+        text = textlist[i]
+        lang = langlist[i]
+        phones, word2ph, norm_text = clean_text_inf(text, lang)
+        bert = get_bert_inf(phones, word2ph, norm_text, lang)
+        bert_list.append(bert)
+    bert = torch.cat(bert_list, dim=1)
+
+    return bert
+
+#i18n("不切"),i18n("凑五句一切"),i18n("凑50字一切"),i18n("按中文句号。切"),i18n("按英文句号.切")
+def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language,how_to_cut=i18n("不切")):
     t0 = ttime()
     prompt_text = prompt_text.strip("\n")
     prompt_language, text = prompt_language, text.strip("\n")
@@ -196,27 +294,38 @@ def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language)
     t1 = ttime()
     prompt_language = dict_language[prompt_language]
     text_language = dict_language[text_language]
-    phones1, word2ph1, norm_text1 = clean_text(prompt_text, prompt_language)
+
-    phones1 = cleaned_text_to_sequence(phones1)
+    if prompt_language == "en":
+        phones1, word2ph1, norm_text1 = clean_text_inf(prompt_text, prompt_language)
+    else:
+        phones1, word2ph1, norm_text1 = nonen_clean_text_inf(prompt_text, prompt_language)
+    if(how_to_cut==i18n("凑五句一切")):text=cut1(text)
+    elif(how_to_cut==i18n("凑50字一切")):text=cut2(text)
+    elif(how_to_cut==i18n("按中文句号。切")):text=cut3(text)
+    elif(how_to_cut==i18n("按英文句号.切")):text=cut4(text)
+    text = text.replace("\n\n","\n").replace("\n\n","\n").replace("\n\n","\n")
+    if(text[-1]not in splits):text+="。"if text_language!="en"else "."
     texts=text.split("\n")
     audio_opt = []
+    if prompt_language == "en":
+        bert1 = get_bert_inf(phones1, word2ph1, norm_text1, prompt_language)
+    else:
+        bert1 = nonen_get_bert_inf(prompt_text, prompt_language)
+    
     for text in texts:
         # 解决输入目标文本的空行导致报错的问题
         if (len(text.strip()) == 0):
             continue
-        phones2, word2ph2, norm_text2 = clean_text(text, text_language)
+        if text_language == "en":
-        phones2 = cleaned_text_to_sequence(phones2)
+            phones2, word2ph2, norm_text2 = clean_text_inf(text, text_language)
-        if prompt_language == "zh":
-            bert1 = get_bert_feature(norm_text1, word2ph1).to(device)
         else:
-            bert1 = torch.zeros(
+            phones2, word2ph2, norm_text2 = nonen_clean_text_inf(text, text_language)
-                (1024, len(phones1)),
+
-                dtype=torch.float16 if is_half == True else torch.float32,
+        if text_language == "en":
-            ).to(device)
+            bert2 = get_bert_inf(phones2, word2ph2, norm_text2, text_language)
-        if text_language == "zh":
-            bert2 = get_bert_feature(norm_text2, word2ph2).to(device)
         else:
-            bert2 = torch.zeros((1024, len(phones2))).to(bert1)
+            bert2 = nonen_get_bert_inf(text, text_language)
+
         bert = torch.cat([bert1, bert2], 1)
 
         all_phoneme_ids = torch.LongTensor(phones1 + phones2).to(device).unsqueeze(0)
@@ -339,6 +448,9 @@ def cut2(inp):
 def cut3(inp):
     inp = inp.strip("\n")
     return "\n".join(["%s。" % item for item in inp.strip("。").split("。")])
+def cut4(inp):
+    inp = inp.strip("\n")
+    return "\n".join(["%s." % item for item in inp.strip(".").split(".")])
 
 def custom_sort_key(s):
     # 使用正则表达式提取字符串中的数字部分和非数字部分
@@ -445,17 +557,24 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
             prompt_language = gr.Dropdown(
                 label=i18n("参考音频的语种"),choices=[i18n("中文"),i18n("英文"),i18n("日文")],value=i18n("中文")
             )
-        gr.Markdown(value=i18n("*请填写需要合成的目标文本"))
+        gr.Markdown(value=i18n("*请填写需要合成的目标文本。中英混合选中文，日英混合选日文，中日混合暂不支持，非目标语言文本自动遗弃。"))
         with gr.Row():
             text = gr.Textbox(label=i18n("需要合成的文本"), value="")
             text_language = gr.Dropdown(
                 label=i18n("需要合成的语种"),choices=[i18n("中文"),i18n("英文"),i18n("日文")],value=i18n("中文")
             )
+            how_to_cut = gr.Radio(
+                label=i18n("怎么切"),
+                choices=[i18n("不切"),i18n("凑五句一切"),i18n("凑50字一切"),i18n("按中文句号。切"),i18n("按英文句号.切"),],
+                value=i18n("凑50字一切"),
+                interactive=True,
+            )
             inference_button = gr.Button(i18n("合成语音"), variant="primary")
             output = gr.Audio(label=i18n("输出的语音"))
+
         inference_button.click(
             get_tts_wav,
-            [inp_ref, prompt_text, prompt_language, text, text_language],
+            [inp_ref, prompt_text, prompt_language, text, text_language,how_to_cut],
             [output],
         )
         history_audio = []
@@ -485,10 +604,12 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
             button1 = gr.Button(i18n("凑五句一切"), variant="primary")
             button2 = gr.Button(i18n("凑50字一切"), variant="primary")
             button3 = gr.Button(i18n("按中文句号。切"), variant="primary")
+            button4 = gr.Button(i18n("按英文句号.切"), variant="primary")
             text_opt = gr.Textbox(label=i18n("切分后文本"), value="")
             button1.click(cut1, [text_inp], [text_opt])
             button2.click(cut2, [text_inp], [text_opt])
             button3.click(cut3, [text_inp], [text_opt])
+            button4.click(cut4, [text_inp], [text_opt])
         gr.Markdown(value=i18n("后续将支持混合语种编码文本输入。"))
 
 app.queue(concurrency_count=511, max_size=1022).launch(

GPT_SoVITS/text/tone_sandhi.py

@@ -455,6 +455,35 @@ class ToneSandhi:
             "电子",
             "人人",
             "虎虎",
+            "幺幺",
+            "干嘛",
+            "学子",
+            "哈哈",
+            "数数",
+            "袅袅",
+            "局地",
+            "以下",
+            "娃哈哈",
+            "花花草草",
+            "留得",
+            "耕地",
+            "想想",
+            "熙熙",
+            "攘攘",
+            "卵子",
+            "死死",
+            "冉冉",
+            "恳恳",
+            "佼佼",
+            "吵吵",
+            "打打",
+            "考考",
+            "整整",
+            "莘莘",
+            "落地",
+            "算子",
+            "家家户户",
+            "青青",
         }
         self.punc = "：，；。？！“”‘’':,;.?!"
 

api.py

@@ -13,7 +13,7 @@
 `-dt` - `默认参考音频文本`
 `-dl` - `默认参考音频语种, "中文","英文","日文","zh","en","ja"`
 
-`-d` - `推理设备, "cuda","cpu"`
+`-d` - `推理设备, "cuda","cpu","mps"`
 `-a` - `绑定地址, 默认"127.0.0.1"`
 `-p` - `绑定端口, 默认9880, 可在 config.py 中指定`
 `-fp` - `覆盖 config.py 使用全精度`
@@ -139,7 +139,6 @@ parser.add_argument("-dt", "--default_refer_text", type=str, default="", help="
 parser.add_argument("-dl", "--default_refer_language", type=str, default="", help="默认参考音频语种")
 
 parser.add_argument("-d", "--device", type=str, default=g_config.infer_device, help="cuda / cpu / mps")
-parser.add_argument("-p", "--port", type=int, default=g_config.api_port, help="default: 9880")
 parser.add_argument("-a", "--bind_addr", type=str, default="127.0.0.1", help="default: 127.0.0.1")
 parser.add_argument("-p", "--port", type=int, default=g_config.api_port, help="default: 9880")
 parser.add_argument("-fp", "--full_precision", action="store_true", default=False, help="覆盖config.is_half为False, 使用全精度")

config.py

@@ -29,6 +29,18 @@ webui_port_subfix = 9871
 
 api_port = 9880
 
+gpu_name = torch.cuda.get_device_name(0)
+if (
+        ("16" in gpu_name and "V100" not in gpu_name.upper())
+        or "P40" in gpu_name.upper()
+        or "P10" in gpu_name.upper()
+        or "1060" in gpu_name
+        or "1070" in gpu_name
+        or "1080" in gpu_name
+):
+    is_half=False
+
+if(infer_device=="cpu"):is_half=False
 
 class Config:
     def __init__(self):

docs/cn/Changelog_CN.md

@@ -29,3 +29,22 @@
 3-优化模型文件排序逻辑
 
 4-中文分词使用jieba_fast代替jieba
+
+### 20240126更新
+
+1-支持输出文本中英混合、日英混合
+
+2-输出可选切分模式
+
+3-修复uvr5读取到目录自动跳出的问题
+
+4-修复多个换行导致推理报错
+
+5-去除推理界面大量冗余log
+
+6-支持mac训练推理
+
+7-自动识别不支持半精度的卡强制单精度。cpu推理下强制单精度。
+
+
+

tools/uvr5/webui.py

@@ -52,7 +52,8 @@ def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format
             paths = [path.name for path in paths]
         for path in paths:
             inp_path = os.path.join(inp_root, path)
-            need_reformat = 1
+            if(os.path.isfile(inp_path)==False):continue
+            try:
                 done = 0
                 try:
                     y, sr = librosa.load(inp_path, sr=None)
@@ -97,6 +98,9 @@ def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format
                             "%s->%s" % (os.path.basename(inp_path), traceback.format_exc())
                         )
                         yield "\n".join(infos)
+            except:
+                infos.append("Oh my god. %s->%s"%(os.path.basename(inp_path), traceback.format_exc()))
+                yield "\n".join(infos)
     except:
         infos.append(traceback.format_exc())
         yield "\n".join(infos)