GPT_SoVITS/inference_webui.py

# Based on GPT-SoVITS-emo by kevinwang676
# I fucking hate this thing. Why does every GPT-SoVITS space have to suck balls?

import os
import torch
from openvoice import se_extractor
from openvoice.api import BaseSpeakerTTS, ToneColorConverter

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

ckpt_base = 'checkpoints/base_speakers/EN'
ckpt_converter = 'checkpoints/converter'
base_speaker_tts = BaseSpeakerTTS(f'{ckpt_base}/config.json', device=device)
base_speaker_tts.load_ckpt(f'{ckpt_base}/checkpoint.pth')

tone_color_converter = ToneColorConverter(f'{ckpt_converter}/config.json', device=device)
tone_color_converter.load_ckpt(f'{ckpt_converter}/checkpoint.pth')

#source_se = torch.load(f'{ckpt_base}/en_default_se.pth').to(device)
#source_se_style = torch.load(f'{ckpt_base}/en_style_se.pth').to(device)

def vc_en(audio_ref, style_mode):
  text = "We have always tried to be at the intersection of technology and liberal arts, to be able to get the best of both, to make extremely advanced products from a technology point of view."
  if style_mode=="default":
    source_se = torch.load(f'{ckpt_base}/en_default_se.pth').to(device)
    reference_speaker = audio_ref
    target_se, audio_name = se_extractor.get_se(reference_speaker, tone_color_converter, target_dir='processed', vad=True)
    save_path = "output.wav"

    # Run the base speaker tts
    src_path = "tmp.wav"
    base_speaker_tts.tts(text, src_path, speaker='default', language='English', speed=1.0)

    # Run the tone color converter
    encode_message = "@MyShell"
    tone_color_converter.convert(
        audio_src_path=src_path,
        src_se=source_se,
        tgt_se=target_se,
        output_path=save_path,
        message=encode_message)

  else:
    source_se = torch.load(f'{ckpt_base}/en_style_se.pth').to(device)
    reference_speaker = audio_ref
    target_se, audio_name = se_extractor.get_se(reference_speaker, tone_color_converter, target_dir='processed', vad=True)

    save_path = "output.wav"

    # Run the base speaker tts
    src_path = "tmp.wav"
    base_speaker_tts.tts(text, src_path, speaker=style_mode, language='English', speed=1.0)

    # Run the tone color converter
    encode_message = "@MyShell"
    tone_color_converter.convert(
        audio_src_path=src_path,
        src_se=source_se,
        tgt_se=target_se,
        output_path=save_path,
        message=encode_message)

  return "output.wav"

# End

import 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
import json

cnhubert_base_path = os.environ.get(
    "cnhubert_base_path", "GPT_SoVITS/pretrained_models/chinese-hubert-base"
)
bert_path = os.environ.get(
    "bert_path", "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large"
)
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
infer_ttswebui = int(infer_ttswebui)
is_share = os.environ.get("is_share", "False")
is_share = eval(is_share)
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")) and torch.cuda.is_available()
import gradio as gr
from transformers import AutoModelForMaskedLM, AutoTokenizer
import numpy as np
import librosa
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。

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)

clm = ""

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


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

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


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_phones_and_bert(text,language):
    if language in {"en","all_zh","all_ja"}:
        language = language.replace("all_","")
        if language == "en":
            LangSegment.setfilters(["en"])
            formattext = " ".join(tmp["text"] for tmp in LangSegment.getTexts(text))
        else:
            # 因无法区别中日文汉字,以用户输入为准
            formattext = text
        while "  " in formattext:
            formattext = formattext.replace("  ", " ")
        phones, word2ph, norm_text = clean_text_inf(formattext, 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)
    elif language in {"zh", "ja","auto"}:
        textlist=[]
        langlist=[]
        LangSegment.setfilters(["zh","ja","en","ko"])
        if language == "auto":
            for tmp in LangSegment.getTexts(text):
                if tmp["lang"] == "ko":
                    langlist.append("zh")
                    textlist.append(tmp["text"])
                else:
                    langlist.append(tmp["lang"])
                    textlist.append(tmp["text"])
        else:
            for tmp in LangSegment.getTexts(text):
                if tmp["lang"] == "en":
                    langlist.append(tmp["lang"])
                else:
                    # 因无法区别中日文汉字,以用户输入为准
                    langlist.append(language)
                textlist.append(tmp["text"])
        print(textlist)
        print(langlist)
        phones_list = []
        bert_list = []
        norm_text_list = []
        for i in range(len(textlist)):
            lang = langlist[i]
            phones, word2ph, norm_text = clean_text_inf(textlist[i], lang)
            bert = get_bert_inf(phones, word2ph, norm_text, lang)
            phones_list.append(phones)
            norm_text_list.append(norm_text)
            bert_list.append(bert)
        bert = torch.cat(bert_list, dim=1)
        phones = sum(phones_list, [])
        norm_text = ''.join(norm_text_list)

    return phones,bert.to(dtype),norm_text


def merge_short_text_in_array(texts, threshold):
    if (len(texts)) < 2:
        return texts
    result = []
    text = ""
    for ele in texts:
        text += ele
        if len(text) >= threshold:
            result.append(text)
            text = ""
    if (len(text) > 0):
        if len(result) == 0:
            result.append(text)
        else:
            result[len(result) - 1] += text
    return result

def get_tts_wav(name, gptmp, svmp, sty, ref_wav_path, prompt_text, prompt_language, text, text_language, how_to_cut="None", top_k=20, top_p=0.6, temperature=0.6, ref_free = False):

    global clm
    if(not ref_wav_path):
        ref_wav_path=f"referenceaudio/{name}/"+referencedata[name][0][sty]
        prompt_text=referencedata[name][1][sty]
    if clm!=name:
        print(f"Switching to model {name}")
        clm=name
        change_gpt_weights(gptmp)
        change_sovits_weights(svmp)
    
    if prompt_text is None or len(prompt_text) == 0:
        ref_free = True
    t0 = ttime()
    prompt_language = dict_language[prompt_language]
    text_language = dict_language[text_language]
    if not ref_free:
        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("Input text:", 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] > 240000 or wav16k.shape[0] < 48000):
            raise OSError("Reference audio too long!!")
        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()

    if (how_to_cut == "4 Sentences"):
        text = cut1(text)
    elif (how_to_cut == "50 Characters"):
        text = cut2(text)
    elif (how_to_cut == "Chinese/Japanese Punctuation"):
        text = cut3(text)
    elif (how_to_cut == "EN Punctuation"):
        text = cut4(text)
    elif (how_to_cut == "All Punctuation"):
        text = cut5(text)
    while "\n\n" in text:
        text = text.replace("\n\n", "\n")
    texts = text.split("\n")
    texts = merge_short_text_in_array(texts, 5)
    audio_opt = []
    if not ref_free:
        phones1,bert1,norm_text1=get_phones_and_bert(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
        if (text[-1] not in splits): text += "。" if text_language != "en" else "."
        phones2,bert2,norm_text2=get_phones_and_bert(text, text_language)
        if not ref_free:
            bert = torch.cat([bert1, bert2], 1)
            all_phoneme_ids = torch.LongTensor(phones1+phones2).to(device).unsqueeze(0)
        else:
            bert = bert2
            all_phoneme_ids = torch.LongTensor(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,
                None if ref_free else prompt,
                bert,
                # prompt_phone_len=ph_offset,
                top_k=top_k,
                top_p=top_p,
                temperature=temperature,
                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)
    mergeitems = ["".join(group) for group in zip(items[::2], items[1::2])]
    # 在句子不存在符号或句尾无符号的时候保证文本完整
    if len(items)%2 == 1:
        mergeitems.append(items[-1])
    opt = "\n".join(mergeitems)
    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 = "GPT_SoVITS/SoVITS_weights"
GPT_weight_root = "GPT_SoVITS/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

def load_models():
    print("Loading models...")
    voices=[]
    ustyles={}
    with open("voicelist.json", "r", encoding="utf-8") as f:
        voc_info = json.load(f)
    for name, info in voc_info.items():
        if not info['enable']:
            continue
        title= info['title']
        gptmodelpath= "%s/%s" % (GPT_weight_root, info['gpt_model_path'])
        sovitsmodelpath= "%s/%s" % (SoVITS_weight_root, info['sovits_model_path'])
        author= info['modelauthor']
        image = info['cover']
        styles = info['styles']
        styletrans = info['styletrans']
        st=[styles, styletrans]
        voices.append((name, title, gptmodelpath, sovitsmodelpath, author, image))
        ustyles[name]=st
        print(f"Indexed model {title}")
    return voices, ustyles

modeldata, referencedata = load_models()

SoVITS_names, GPT_names = get_weights_names()

#print(os.getcwd())
#for r, _, f in os.walk(os.getcwd()):
#    for n in f:
#        print(os.path.join(r, n))

#Gradio preload
text = gr.TextArea(label="Input Text", value="Hello there! This is test audio of a new text to speech tool.")
text_language = gr.Dropdown(label="Language", choices=["EN", "JP", "ZH", "ZH/EN", "JP/EN", "Automatic"], value="EN")
how_to_cut = gr.Dropdown(label="Slicing Method",
    choices=["None", "4 Sentences", "50 Characters", "ZH/JP Punctuation", "EN Punctuation", "All Punctuation" ],
    value="4 Sentences",
    interactive=True,
)
top_k = gr.Slider(minimum=1,maximum=100,step=1,label="top_k",value=5,interactive=True)
top_p = gr.Slider(minimum=0,maximum=1,step=0.05,label="top_p",value=1,interactive=True)
temperature = gr.Slider(minimum=0,maximum=1,step=0.05,label="temperature",value=1,interactive=True)

#Main gradio
with gr.Blocks(title="Lemonfoot GPT-SoVITS") as app:
    gr.Markdown(
        "# Lemonfoot GPT-SoVITS 🚀🍋\n"
        "### Space by Kit Lemonfoot / Noel Shirogane's High Flying Birds\n"
        "Based on code originally by RVC_Boss and kevinwang676\n\n"
        "Do no evil.\n\n"
        "**NOTE:** *This is more or less a test Space*. HuggingFace Spaces are not capable of running GPT-SoVITS efficiently; a single generation may take upwards of an hour to infer one sentence. "
        "If you wish to use these models for legitimate generation, it is recommended to [download the models individually](https://huggingface.co/Kit-Lemonfoot/kitlemonfoot_gptsovits_models) and run GPT-SoVITS locally."
    )
    for (name, title, gptmodelpath, sovitsmodelpath, author, image) in modeldata:
        with gr.TabItem(name):
            with gr.Row():
                with gr.Column():
                    n = gr.Textbox(value=name, visible=False, interactive=False)
                    gptmp = gr.Textbox(value=gptmodelpath, visible=False, interactive=False)
                    svmp = gr.Textbox(value=sovitsmodelpath, visible=False, interactive=False)
                    gr.Markdown(f"**{title}**\n\n Dataset author: {author}")
                    gr.Image(f"images/{image}", label=None, show_label=False, width=300, show_download_button=False, container=False, show_share_button=False)
                with gr.Column():
                    with gr.TabItem("Style using a preset"):
                        sty = gr.Dropdown(
                            label="Current style",
                            choices=referencedata[name][0].keys(),
                            value="Neutral",
                            interactive=True
                        )
                    with gr.TabItem("Style using a different audio"):
                        with gr.Column():
                            ref_audio_path = gr.Audio(label="Reference Audio", type="filepath")
                            ref_text_free = gr.Checkbox(label="Enables no text-reference mode.", value=False, interactive=True)
                            prompt_text = gr.Textbox(label="Reference Audio Text", interactive=True)
                            prompt_language = gr.Textbox(value="EN", visible=False, interactive=False)
                with gr.Column():
                    inference_button = gr.Button("Synthesize", variant="primary")
                    output = gr.Audio(label="Output")

                    inference_button.click(
                        get_tts_wav,
                        inputs=[n, gptmp, svmp, sty, ref_audio_path, prompt_text, prompt_language, text, text_language, how_to_cut, top_k, top_p, temperature, ref_text_free],
                        outputs=[output]
                    )

    #bottom info
    with gr.Row():
        with gr.Column():
            text.render()
            text_language.render()
            how_to_cut.render()
        with gr.Column():
            gr.Markdown("### GPT Sampling Parameters")
            top_k.render()
            top_p.render()
            temperature.render()

app.queue().launch()