app.py

'''

按中英混合识别

按日英混合识别

多语种启动切分识别语种

全部按中文识别

全部按英文识别

全部按日文识别

'''
import os, re, logging
import LangSegment
logging.getLogger("markdown_it").setLevel(logging.ERROR)
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

gpt_path = os.environ.get(
    "gpt_path", "models/skd/xxx-e15.ckpt"
)

sovits_path = os.environ.get(
    "sovits_path", "models/skd/xxx_e12_s108.pth"
)

cnhubert_base_path = os.environ.get(
    "cnhubert_base_path", "pretrained_models/chinese-hubert-base"
)
bert_path = os.environ.get(
    "bert_path", "pretrained_models/chinese-roberta-wwm-ext-large"
)
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
infer_ttswebui = int(infer_ttswebui)
if "_CUDA_VISIBLE_DEVICES" in os.environ:
    os.environ["CUDA_VISIBLE_DEVICES"] = os.environ["_CUDA_VISIBLE_DEVICES"]
is_half = eval(os.environ.get("is_half", "True"))
import gradio as gr
from transformers import AutoModelForMaskedLM, AutoTokenizer
import numpy as np
import librosa, torch
from feature_extractor import cnhubert

cnhubert.cnhubert_base_path = cnhubert_base_path

from module.models import SynthesizerTrn
from AR.models.t2s_lightning_module import Text2SemanticLightningModule
from text import cleaned_text_to_sequence
from text.cleaner import clean_text
from time import time as ttime
from module.mel_processing import spectrogram_torch
from my_utils import load_audio

os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1'  # 确保直接启动推理UI时也能够设置。

if torch.cuda.is_available():
    device = "cuda"
elif torch.backends.mps.is_available():
    device = "mps"
else:
    device = "cpu"

tokenizer = AutoTokenizer.from_pretrained(bert_path)
bert_model = AutoModelForMaskedLM.from_pretrained(bert_path)
if is_half == True:
    bert_model = bert_model.half().to(device)
else:
    bert_model = bert_model.to(device)


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)
        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)
    phone_level_feature = []
    for i in range(len(word2ph)):
        repeat_feature = res[i].repeat(word2ph[i], 1)
        phone_level_feature.append(repeat_feature)
    phone_level_feature = torch.cat(phone_level_feature, dim=0)
    return phone_level_feature.T


class DictToAttrRecursive(dict):
    def __init__(self, input_dict):
        super().__init__(input_dict)
        for key, value in input_dict.items():
            if isinstance(value, dict):
                value = DictToAttrRecursive(value)
            self[key] = value
            setattr(self, key, value)

    def __getattr__(self, item):
        try:
            return self[item]
        except KeyError:
            raise AttributeError(f"Attribute {item} not found")

    def __setattr__(self, key, value):
        if isinstance(value, dict):
            value = DictToAttrRecursive(value)
        super(DictToAttrRecursive, self).__setitem__(key, value)
        super().__setattr__(key, value)

    def __delattr__(self, item):
        try:
            del self[item]
        except KeyError:
            raise AttributeError(f"Attribute {item} not found")


ssl_model = cnhubert.get_model()
if is_half == True:
    ssl_model = ssl_model.half().to(device)
else:
    ssl_model = ssl_model.to(device)


def change_sovits_weights(sovits_path):
    global vq_model, hps
    dict_s2 = torch.load(sovits_path, map_location="cpu")
    hps = dict_s2["config"]
    hps = DictToAttrRecursive(hps)
    hps.model.semantic_frame_rate = "25hz"
    vq_model = SynthesizerTrn(
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        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)
    else:
        vq_model = vq_model.to(device)
    vq_model.eval()
    print(vq_model.load_state_dict(dict_s2["weight"], strict=False))


change_sovits_weights(sovits_path)


def change_gpt_weights(gpt_path):
    global hz, max_sec, t2s_model, config
    hz = 50
    dict_s1 = torch.load(gpt_path, map_location="cpu")
    config = dict_s1["config"]
    max_sec = config["data"]["max_sec"]
    t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
    t2s_model.load_state_dict(dict_s1["weight"])
    if is_half == True:
        t2s_model = t2s_model.half()
    t2s_model = t2s_model.to(device)
    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):
    audio = load_audio(filename, int(hps.data.sampling_rate))
    audio = torch.FloatTensor(audio)
    audio_norm = audio
    audio_norm = audio_norm.unsqueeze(0)
    spec = spectrogram_torch(
        audio_norm,
        hps.data.filter_length,
        hps.data.sampling_rate,
        hps.data.hop_length,
        hps.data.win_length,
        center=False,
    )
    return spec


dict_language = {
    ("中文"): "all_zh",#全部按中文识别
    ("英文"): "en",#全部按英文识别#######不变
    ("日文"): "all_ja",#全部按日文识别
    ("中英混合"): "zh",#按中英混合识别####不变
    ("日英混合"): "ja",#按日英混合识别####不变
    ("多语种混合"): "auto",#多语种启动切分识别语种
}


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)
    # Merge punctuation into previous word
    for i in range(len(textlist)-1, 0, -1):
        if re.match(r'^[\W_]+$', textlist[i]):
            textlist[i-1] += textlist[i]
            del textlist[i]
            del langlist[i]
    # Merge consecutive words with the same language tag
    i = 0
    while i < len(langlist) - 1:
        if langlist[i] == langlist[i+1]:
            textlist[i] += textlist[i+1]
            del textlist[i+1]
            del langlist[i+1]
        else:
            i += 1

    return textlist, langlist


def clean_text_inf(text, language):
    phones, word2ph, norm_text = clean_text(text, language.replace("all_",""))
    phones = cleaned_text_to_sequence(phones)
    return phones, word2ph, norm_text

dtype=torch.float16 if is_half == True else torch.float32
def get_bert_inf(phones, word2ph, norm_text, language):
    language=language.replace("all_","")
    if language == "zh":
        bert = get_bert_feature(norm_text, word2ph).to(device)#.to(dtype)
    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):
    if(language!="auto"):
        textlist, langlist = splite_en_inf(text, language)
    else:
        textlist=[]
        langlist=[]
        for tmp in LangSegment.getTexts(text):
            langlist.append(tmp["lang"])
            textlist.append(tmp["text"])
    print(textlist)
    print(langlist)
    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 == "zh":
            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):
    if(language!="auto"):
        textlist, langlist = splite_en_inf(text, language)
    else:
        textlist=[]
        langlist=[]
        for tmp in LangSegment.getTexts(text):
            langlist.append(tmp["lang"])
            textlist.append(tmp["text"])
    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


splits = {"，", "。", "？", "！", ",", ".", "?", "!", "~", ":", "：", "—", "…", }


def get_first(text):
    pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]"
    text = re.split(pattern, text)[0].strip()
    return text


def get_cleaned_text_final(text,language):
    if language in {"en","all_zh","all_ja"}:
        phones, word2ph, norm_text = clean_text_inf(text, language)
    elif language in {"zh", "ja","auto"}:
        phones, word2ph, norm_text = nonen_clean_text_inf(text, language)
    return phones, word2ph, norm_text

def get_bert_final(phones, word2ph, text,language,device):
    if language == "en":
        bert = get_bert_inf(phones, word2ph, text, language)
    elif language in {"zh", "ja","auto"}:
        bert = nonen_get_bert_inf(text, language)
    elif language == "all_zh":
        bert = get_bert_feature(text, word2ph).to(device)
    else:
        bert = torch.zeros((1024, len(phones))).to(device)
    return bert

def get_tts_wav(ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut=("不切")):
    t0 = ttime()
    prompt_language = dict_language[prompt_language]
    text_language = dict_language[text_language]
    prompt_text = prompt_text.strip("\n")
    if (prompt_text[-1] not in splits): prompt_text += "。" if prompt_language != "en" else "."
    text = text.strip("\n")
    if (text[0] not in splits and len(get_first(text)) < 4): text = "。" + text if text_language != "en" else "." + text
    print(("实际输入的参考文本:"), prompt_text)
    print(("实际输入的目标文本:"), text)
    zero_wav = np.zeros(
        int(hps.data.sampling_rate * 0.3),
        dtype=np.float16 if is_half == True else np.float32,
    )
    with torch.no_grad():
        wav16k, sr = librosa.load(ref_wav_path, sr=16000)
        if (wav16k.shape[0] > 160000 or wav16k.shape[0] < 48000):
            raise OSError(("参考音频在3~10秒范围外，请更换！"))
        wav16k = torch.from_numpy(wav16k)
        zero_wav_torch = torch.from_numpy(zero_wav)
        if is_half == True:
            wav16k = wav16k.half().to(device)
            zero_wav_torch = zero_wav_torch.half().to(device)
        else:
            wav16k = wav16k.to(device)
            zero_wav_torch = zero_wav_torch.to(device)
        wav16k = torch.cat([wav16k, zero_wav_torch])
        ssl_content = ssl_model.model(wav16k.unsqueeze(0))[
            "last_hidden_state"
        ].transpose(
            1, 2
        )  # .float()
        codes = vq_model.extract_latent(ssl_content)
        prompt_semantic = codes[0, 0]
    t1 = ttime()

    phones1, word2ph1, norm_text1=get_cleaned_text_final(prompt_text, prompt_language)

    if (how_to_cut == ("凑四句一切")):
        text = cut1(text)
    elif (how_to_cut == ("凑50字一切")):
        text = cut2(text)
    elif (how_to_cut == ("按中文句号。切")):
        text = cut3(text)
    elif (how_to_cut == ("按英文句号.切")):
        text = cut4(text)
    elif (how_to_cut == ("按标点符号切")):
        text = cut5(text)
    text = text.replace("\n\n", "\n").replace("\n\n", "\n").replace("\n\n", "\n")
    print(("实际输入的目标文本(切句后):"), text)
    texts = text.split("\n")
    audio_opt = []
    bert1=get_bert_final(phones1, word2ph1, norm_text1,prompt_language,device).to(dtype)

    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
        if (text[-1] not in splits): text += "。" if text_language != "en" else "."
        print(("实际输入的目标文本(每句):"), text)
        phones2, word2ph2, norm_text2 = get_cleaned_text_final(text, text_language)
        bert2 = get_bert_final(phones2, word2ph2, norm_text2, text_language, device).to(dtype)
        print(bert1,bert2)
        bert = torch.cat([bert1, bert2], 1)

        all_phoneme_ids = torch.LongTensor(phones1 + phones2).to(device).unsqueeze(0)
        bert = bert.to(device).unsqueeze(0)
        all_phoneme_len = torch.tensor([all_phoneme_ids.shape[-1]]).to(device)
        prompt = prompt_semantic.unsqueeze(0).to(device)
        t2 = ttime()
        with torch.no_grad():
            # pred_semantic = t2s_model.model.infer(
            pred_semantic, idx = t2s_model.model.infer_panel(
                all_phoneme_ids,
                all_phoneme_len,
                prompt,
                bert,
                # prompt_phone_len=ph_offset,
                top_k=config["inference"]["top_k"],
                early_stop_num=hz * max_sec,
            )
        t3 = ttime()
        # print(pred_semantic.shape,idx)
        pred_semantic = pred_semantic[:, -idx:].unsqueeze(
            0
        )  # .unsqueeze(0)#mq要多unsqueeze一次
        refer = get_spepc(hps, ref_wav_path)  # .to(device)
        if is_half == True:
            refer = refer.half().to(device)
        else:
            refer = refer.to(device)
        # audio = vq_model.decode(pred_semantic, all_phoneme_ids, refer).detach().cpu().numpy()[0, 0]
        audio = (
            vq_model.decode(
                pred_semantic, torch.LongTensor(phones2).to(device).unsqueeze(0), refer
            )
                .detach()
                .cpu()
                .numpy()[0, 0]
        )  ###试试重建不带上prompt部分
        max_audio=np.abs(audio).max()#简单防止16bit爆音
        if max_audio>1:audio/=max_audio
        audio_opt.append(audio)
        audio_opt.append(zero_wav)
        t4 = ttime()
    print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
    yield hps.data.sampling_rate, (np.concatenate(audio_opt, 0) * 32768).astype(
        np.int16
    )


def split(todo_text):
    todo_text = todo_text.replace("……", "。").replace("——", "，")
    if todo_text[-1] not in splits:
        todo_text += "。"
    i_split_head = i_split_tail = 0
    len_text = len(todo_text)
    todo_texts = []
    while 1:
        if i_split_head >= len_text:
            break  # 结尾一定有标点，所以直接跳出即可，最后一段在上次已加入
        if todo_text[i_split_head] in splits:
            i_split_head += 1
            todo_texts.append(todo_text[i_split_tail:i_split_head])
            i_split_tail = i_split_head
        else:
            i_split_head += 1
    return todo_texts


def cut1(inp):
    inp = inp.strip("\n")
    inps = split(inp)
    split_idx = list(range(0, len(inps), 4))
    split_idx[-1] = None
    if len(split_idx) > 1:
        opts = []
        for idx in range(len(split_idx) - 1):
            opts.append("".join(inps[split_idx[idx]: split_idx[idx + 1]]))
    else:
        opts = [inp]
    return "\n".join(opts)


def cut2(inp):
    inp = inp.strip("\n")
    inps = split(inp)
    if len(inps) < 2:
        return inp
    opts = []
    summ = 0
    tmp_str = ""
    for i in range(len(inps)):
        summ += len(inps[i])
        tmp_str += inps[i]
        if summ > 50:
            summ = 0
            opts.append(tmp_str)
            tmp_str = ""
    if tmp_str != "":
        opts.append(tmp_str)
    # print(opts)
    if len(opts) > 1 and len(opts[-1]) < 50:  ##如果最后一个太短了，和前一个合一起
        opts[-2] = opts[-2] + opts[-1]
        opts = opts[:-1]
    return "\n".join(opts)


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(".")])


# contributed by https://github.com/AI-Hobbyist/GPT-SoVITS/blob/main/GPT_SoVITS/inference_webui.py
def cut5(inp):
    # if not re.search(r'[^\w\s]', inp[-1]):
    # inp += '。'
    inp = inp.strip("\n")
    punds = r'[,.;?!、，。？！;：]'
    items = re.split(f'({punds})', inp)
    items = ["".join(group) for group in zip(items[::2], items[1::2])]
    opt = "\n".join(items)
    return opt


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
    parts = re.split('(\d+)', s)
    # 将数字部分转换为整数，非数字部分保持不变
    parts = [int(part) if part.isdigit() else part for part in parts]
    return parts


def change_choices():
    SoVITS_names, GPT_names = get_weights_names()
    return {"choices": sorted(SoVITS_names, key=custom_sort_key), "__type__": "update"}, {"choices": sorted(GPT_names, key=custom_sort_key), "__type__": "update"}


pretrained_sovits_name = "GPT_SoVITS/pretrained_models/s2G488k.pth"
pretrained_gpt_name = "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt"
SoVITS_weight_root = "SoVITS_weights"
GPT_weight_root = "GPT_weights"
os.makedirs(SoVITS_weight_root, exist_ok=True)
os.makedirs(GPT_weight_root, exist_ok=True)


def get_weights_names():
    SoVITS_names = [pretrained_sovits_name]
    for name in os.listdir(SoVITS_weight_root):
        if name.endswith(".pth"): SoVITS_names.append("%s/%s" % (SoVITS_weight_root, name))
    GPT_names = [pretrained_gpt_name]
    for name in os.listdir(GPT_weight_root):
        if name.endswith(".ckpt"): GPT_names.append("%s/%s" % (GPT_weight_root, name))
    return SoVITS_names, GPT_names


SoVITS_names, GPT_names = get_weights_names()

with gr.Blocks(title="GPT-SoVITS WebUI") as app:
    gr.Markdown(
        value=("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责. <br>如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录<b>LICENSE</b>.")
    )
    with gr.Group():
        gr.Markdown(value=("模型切换"))
        with gr.Row():
            GPT_dropdown = gr.Dropdown(label=("GPT模型列表"), choices=sorted(GPT_names, key=custom_sort_key), value=gpt_path, interactive=True)
            SoVITS_dropdown = gr.Dropdown(label=("SoVITS模型列表"), choices=sorted(SoVITS_names, key=custom_sort_key), value=sovits_path, interactive=True)
            refresh_button = gr.Button(("刷新模型路径"), variant="primary")
            refresh_button.click(fn=change_choices, inputs=[], outputs=[SoVITS_dropdown, GPT_dropdown])
            SoVITS_dropdown.change(change_sovits_weights, [SoVITS_dropdown], [])
            GPT_dropdown.change(change_gpt_weights, [GPT_dropdown], [])
        gr.Markdown(value=("*请上传并填写参考信息"))
        with gr.Row():
            inp_ref = gr.Audio(label=("请上传3~10秒内参考音频，超过会报错！"), type="filepath")
            prompt_text = gr.Textbox(label=("参考音频的文本"), value="")
            prompt_language = gr.Dropdown(
                label=("参考音频的语种"), choices=[("中文"), ("英文"), ("日文"), ("中英混合"), ("日英混合"), ("多语种混合")], value=("中文")
            )
        gr.Markdown(value=("*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=("需要合成的文本"), value="")
            text_language = gr.Dropdown(
                label=("需要合成的语种"), choices=[("中文"), ("英文"), ("日文"), ("中英混合"), ("日英混合"), ("多语种混合")], value=("中文")
            )
            how_to_cut = gr.Radio(
                label=("怎么切"),
                choices=[("不切"), ("凑四句一切"), ("凑50字一切"), ("按中文句号。切"), ("按英文句号.切"), ("按标点符号切"), ],
                value=("凑四句一切"),
                interactive=True,
            )
            inference_button = gr.Button(("合成语音"), variant="primary")
            output = gr.Audio(label=("输出的语音"))

        inference_button.click(
            get_tts_wav,
            [inp_ref, prompt_text, prompt_language, text, text_language, how_to_cut],
            [output],
        )

        gr.Markdown(value=("文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。"))
        with gr.Row():
            text_inp = gr.Textbox(label=("需要合成的切分前文本"), value="")
            button1 = gr.Button(("凑四句一切"), variant="primary")
            button2 = gr.Button(("凑50字一切"), variant="primary")
            button3 = gr.Button(("按中文句号。切"), variant="primary")
            button4 = gr.Button(("按英文句号.切"), variant="primary")
            button5 = gr.Button(("按标点符号切"), variant="primary")
            text_opt = gr.Textbox(label=("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])
            button5.click(cut5, [text_inp], [text_opt])
        gr.Markdown(value=("后续将支持混合语种编码文本输入。"))

app.queue(concurrency_count=511, max_size=1022).launch(inbrowser=True)