app.py

import os
import traceback
import gradio as gr
import numpy as np
import torch
import spaces
from loguru import logger
from transformers import AutoTokenizer, AutoModelForCausalLM
from funasr_onnx import Paraformer
from huggingface_hub import snapshot_download

from tools.wer import compute_wers

os.environ["EINX_FILTER_TRACEBACK"] = "false"

from i18n import i18n
from text.chn_text_norm.text import Text as ChnNormedText
from xcodec2.modeling_xcodec2 import XCodec2Model


TEXTBOX_PLACEHOLDER = i18n("Put your text here.")

# ===== Hugging Face Model IDs =====
LLASA_MODEL_ID = "ASLP-lab/VoiceSculptor-VD"
XCODEC_MODEL_ID = "HKUSTAudio/xcodec2"
PARAFORMER_REPO_ID = "funasr/Paraformer-large"

# logo
LOGO_URL = "https://raw.githubusercontent.com/ASLP-lab/VoiceSculptor/main/assets/logo.png"

# ===== Global cache =====
model = None
codec_model = None
asr_model = None
tokenizer = None
device= 'cuda'

def normalize_text_final(user_input: str) -> str:
    return ChnNormedText(raw_text=user_input).normalize()


def extract_speech_ids(token_strs: list[str]) -> list[int]:
    """把 tokenizer 输出的 token 字符串列表中形如 <|s_123|> 的 token 提取成 int id"""
    speech_ids = []
    for t in token_strs:
        if t.startswith("<|s_") and t.endswith("|>"):
            num_str = t[4:-2]
            try:
                speech_ids.append(int(num_str))
            except Exception:
                logger.warning(f"Bad speech token: {t}")
    return speech_ids


def get_asr(asr_model: Paraformer, wav_list: list[np.ndarray]) -> list[str]:
    """wav_list: list of 1D numpy waveform (16k)"""
    try:
        result = asr_model(wav_list)
        if isinstance(result, dict):
            result = [result]

        texts = []
        for res in result:
            preds = res.get("preds", None)
            if preds is None:
                texts.append(res.get("text", ""))
            else:
                texts.append(preds[0] if len(preds) > 0 else "")

        if len(texts) != len(wav_list):
            logger.warning(f"[ASR] batch返回数量不一致: got {len(texts)} expect {len(wav_list)}，fallback逐条补齐")
            texts = []
            for w in wav_list:
                try:
                    r = asr_model(w)
                    if isinstance(r, list) and len(r) > 0:
                        r0 = r[0]
                        preds = r0.get("preds", None)
                        texts.append(preds[0] if preds else r0.get("text", ""))
                    elif isinstance(r, dict):
                        preds = r.get("preds", None)
                        texts.append(preds[0] if preds else r.get("text", ""))
                    else:
                        texts.append("")
                except Exception:
                    texts.append("")
        return texts

    except Exception as e:
        logger.warning(f"[ASR] batch失败，fallback逐条: {e}")
        texts = []
        for w in wav_list:
            try:
                r = asr_model(w)
                if isinstance(r, list) and len(r) > 0:
                    r0 = r[0]
                    preds = r0.get("preds", None)
                    texts.append(preds[0] if preds else r0.get("text", ""))
                elif isinstance(r, dict):
                    preds = r.get("preds", None)
                    texts.append(preds[0] if preds else r.get("text", ""))
                else:
                    texts.append("")
            except Exception:
                texts.append("")
        return texts


def _safe_load_tokenizer(model_id: str) -> AutoTokenizer:
    try:
        tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
    except TypeError:
        tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True, use_fast=False)

    if tok.pad_token_id is None:
        if tok.eos_token_id is not None:
            tok.pad_token = tok.eos_token
    return tok


def _safe_load_lm(model_id: str, device: str) -> AutoModelForCausalLM:
    m = AutoModelForCausalLM.from_pretrained(
        model_id,
        trust_remote_code=True,
    )
    m.eval().to(device)
    return m

def load_models(force_device: str | None = None):
    global model, codec_model, asr_model, tokenizer

    logger.info(f"Using device: {device}")

    if tokenizer is None:
        logger.info("Loading tokenizer...")
        tokenizer = _safe_load_tokenizer(LLASA_MODEL_ID)

    if model is None:
        logger.info("Loading AutoModelForCausalLM...")
        model = _safe_load_lm(LLASA_MODEL_ID, device=device)

    if codec_model is None:
        logger.info("Loading XCodec2...")
        codec_model = XCodec2Model.from_pretrained(XCODEC_MODEL_ID).eval().to(device)

    if asr_model is None:
        logger.info("Loading Paraformer (funasr_onnx)...")
        paraformer_dir = snapshot_download(
            repo_id=PARAFORMER_REPO_ID,
            local_dir="checkpoints/Paraformer-large",
            local_dir_use_symlinks=False,
        )
        asr_model = Paraformer(paraformer_dir, batch_size=5, quantize=True)

    logger.info("All models loaded.")

load_models()

@spaces.GPU
def inference_batch_transformers(
    lm: AutoModelForCausalLM,
    codec: XCodec2Model,
    tok: AutoTokenizer,
    refined_text: str,
    instruct_text: str,
    control_tags: str,
    batch_size: int = 5,
    max_new_tokens: int = 2048,
) -> list[tuple[int, np.ndarray]]:

    refined_text_norm = normalize_text_final(refined_text)
    instruct_text_norm = normalize_text_final(instruct_text)

    if len(refined_text_norm) < 5:
        raise ValueError("输入文本长度不能少于5个字符")
    if len(refined_text_norm) > 150:
        raise ValueError("输入文本长度不能超过150个字符")

    target_text = instruct_text_norm + "<|endofprompt|>" + control_tags + refined_text_norm
    formatted_text = f"<|TEXT_UNDERSTANDING_START|>{target_text}<|TEXT_UNDERSTANDING_END|>"
    chat = [
        {"role": "user", "content": "Convert the text to speech:" + formatted_text},
        {"role": "assistant", "content": "<|SPEECH_GENERATION_START|>"},
    ]

    input_ids_1 = tok.apply_chat_template(
        chat,
        tokenize=True,
        return_tensors="pt",
        continue_final_message=True,
    ).to(device)

    speech_end_id = tok.convert_tokens_to_ids("<|SPEECH_GENERATION_END|>")

    outputs = lm.generate(
        input_ids=input_ids_1,
        do_sample=True,
        top_p=0.95,
        temperature=0.9,
        top_k=15,
        repetition_penalty=1.05,
        max_new_tokens=max_new_tokens,
        eos_token_id=speech_end_id,
        num_return_sequences=batch_size, 
        use_cache=True,
    )

    prompt_len = input_ids_1.shape[1]
    audios: list[tuple[int, np.ndarray]] = []

    for i in range(outputs.shape[0]):
        gen_ids = outputs[i, prompt_len:].tolist()
        if len(gen_ids) > 0 and gen_ids[-1] == speech_end_id:
            gen_ids = gen_ids[:-1]

        token_strs = tok.convert_ids_to_tokens(gen_ids, skip_special_tokens=False)
        speech_ids = extract_speech_ids(token_strs)

        if len(speech_ids) == 0:
            logger.warning("[TTS] No speech tokens extracted, outputting silence.")
            audios.append((16000, np.zeros((16000,), dtype=np.float32)))
            continue

        speech_tokens_t = torch.tensor(speech_ids, device=device).unsqueeze(0).unsqueeze(0)
        wav = codec.decode_code(speech_tokens_t)
        wav = wav.squeeze(0).squeeze(0).detach().cpu().numpy().astype(np.float32)
        audios.append((16000, wav))

    return audios


def build_control_tags(age, gender, pitch, pitch_var, volume, speed, emo):
    tag_map = {
        "小孩": "<|小孩|>", "青年": "<|青年|>", "中年": "<|中年|>", "老年": "<|老年|>",
        "男性": "<|男性|>", "女性": "<|女性|>",
        "音调很高": "<|音调很高|>", "音调较高": "<|音调较高|>", "音调中等": "<|音调中等|>",
        "音调较低": "<|音调较低|>", "音调很低": "<|音调很低|>",
        "音调变化很强": "<|音调变化很强|>", "音调变化较强": "<|音调变化较强|>", "音调变化一般": "<|音调变化一般|>",
        "音调变化较弱": "<|音调变化较弱|>", "音调变化很弱": "<|音调变化很弱|>",
        "音量很大": "<|音量很大|>", "音量较大": "<|音量较大|>", "音量中等": "<|音量中等|>",
        "音量较小": "<|音量较小|>", "音量很小": "<|音量很小|>",
        "语速很快": "<|语速很快|>", "语速较快": "<|语速较快|>", "语速中等": "<|语速中等|>",
        "语速较慢": "<|语速较慢|>", "语速很慢": "<|语速很慢|>",
        "开心": "<|开心|>", "生气": "<|生气|>", "难过": "<|难过|>", "惊讶": "<|惊讶|>",
        "厌恶": "<|厌恶|>", "害怕": "<|害怕|>",
    }
    tags = []
    for v in [gender, age, speed, volume, pitch, pitch_var, emo]:
        if v != "不指定":
            tags.append(tag_map[v])
    return "".join(tags)

@spaces.GPU
def inference_select_best3(refined_text, instruct_text, age, gender, pitch, pitch_var, volume, speed, emo):
    
    control_tags = build_control_tags(age, gender, pitch, pitch_var, volume, speed, emo)

    try:
        audios5 = inference_batch_transformers(
            lm=model,
            codec=codec_model,
            tok=tokenizer,
            refined_text=refined_text,
            instruct_text=instruct_text,
            control_tags=control_tags,
            batch_size=5,
            max_new_tokens=2048,
        )

        wav_list = [wav for (_, wav) in audios5]
        asr_texts = get_asr(asr_model, wav_list)

        refined_text_norm = normalize_text_final(refined_text)
        gt_texts = [refined_text_norm] * len(asr_texts)
        wers = compute_wers(gt_texts, asr_texts, lang="zh")

        for i, (hyp, w) in enumerate(zip(asr_texts, wers)):
            logger.info(f"[ASR/WER] idx={i} wer={w:.4f}  gt='{refined_text_norm}'  asr='{hyp}'")

        best_idx = np.argsort(np.array(wers))[:3].tolist()
        best3 = [audios5[i] for i in best_idx]
        return best3[0], best3[1], best3[2]

    except Exception as e:
        logger.error(f"推理/ASR/WER 失败: {e}", exc_info=True)
        logger.error("错误详细信息:\n" + traceback.format_exc())
        return None, None, None


def build_app():

    INSTRUCT_TEMPLATES = {
        "自定义": "",
        "default": "这是一位男性评书表演者，用传统说唱腔调，以变速节奏和韵律感极强的语速讲述江湖故事，音量时高时低，充满江湖气。",
        "幼儿园女教师-温柔甜美": "这是一位幼儿园女教师，用甜美明亮的嗓音，以极慢且富有耐心的语速，带着温柔鼓励的情感，用标准普通话给小朋友讲睡前故事，音量轻柔适中，咬字格外清晰。",
        "电台主播-平静温柔": "深夜电台主播，男性、音调偏低、语速偏慢、音量小；情绪平静带点忧伤，语气温柔；音色微哑",
        "成熟御姐-温柔暧昧": "成熟御姐风格，语速偏慢，音量适中，情绪慵懒暧昧，语气温柔笃定带掌控感，磁性低音，吐字清晰，尾音微挑，整体有贴近感与撩人的诱惑。",
        "年轻妈妈-温暖安抚": "年轻妈妈哄孩子入睡，女性、音调柔和偏低、语速偏慢、音量偏小但清晰；情绪温暖安抚、充满耐心与爱意，语气轻柔哄劝、像贴近耳边低声说话；音色软糯，吐字清晰、节奏舒缓。",
        "小女孩-尖锐清脆": "一位7岁的小女孩，用天真高亢的童声，以不稳定的快节奏，充满兴奋和炫耀地背诵乘法口诀，音调忽高忽低，带着儿童特有的尖锐清脆。",
        "老奶奶-沙哑低沉": "一位慈祥的老奶奶，用沙哑低沉的嗓音，以极慢而温暖的语速讲述民间传说，音量微弱但清晰，带着怀旧和神秘的情感。",
        "诗歌朗诵-雄浑有力": "一位男性现代诗朗诵者，用深沉磁性的低音，以顿挫有力的节奏演绎艾青诗歌，音量洪亮，情感激昂澎湃。",
        "童话风格-甜美夸张": "这是一位女性童话旁白朗诵者，用甜美夸张的童声，以跳跃变化的语速讲述《安徒生童话》，音调偏高，充满奇幻色彩。",
        "评书风格-抑扬顿挫": "这是一位男性评书表演者，用传统说唱腔调，以变速节奏和韵律感极强的语速讲述江湖故事，音量时高时低，充满江湖气。",
        "新闻风格-平静专业": "这是一位女性新闻主播，用标准普通话以清晰明亮的中高音，以平稳专业的语速播报时事新闻，音量洪亮，情感客观中立。",
        "相声风格-夸张幽默": "这是一位男性相声表演者，用夸张幽默的嗓音，以时快时慢的节奏抖包袱，音调起伏大，充满喜感和节奏感。",
        "悬疑小说-低沉神秘": "一位男性悬疑小说演播者，用低沉神秘的嗓音，以时快时慢的变速节奏营造紧张氛围，音量忽高忽低，充满悬念感。",
        "戏剧表演-夸张戏剧": "这是一位男性戏剧表演者，用夸张戏剧化的嗓音，以忽高忽低的音调和时快时慢的语速表演独白，充满张力。",
        "法治节目-庄严庄重": "这是一位男性法治节目主持人，用严肃庄重的嗓音，以平稳有力的语速讲述案件，音量适中，体现法律的威严。",
        "纪录片旁白-低沉磁性": "这是一位男性纪录片旁白，用深沉磁性的嗓音，以缓慢而富有画面感的语速讲述自然奇观，音量适中，充满敬畏和诗意。",
        "广告配音-沧桑浑厚": "这是一位男性白酒品牌广告配音，用沧桑浑厚的嗓音，以缓慢而豪迈的语速，音量洪亮，传递历史底蕴和男人情怀。",
        "冥想引导师-空灵悠长": "一位女性冥想引导师，用空灵悠长的气声，以极慢而飘渺的语速，配合环境音效，音量轻柔，营造禅意空间。",
        "ASMR-气声耳语": "一位女性ASMR主播，用气声耳语，以极慢而细腻的语速，配合唇舌音，音量极轻，营造极度放松的氛围。",
    }

    TEXT_REQUIREMENTS = {
        "自定义": "",
        "default": "话说那武松，提着哨棒，直奔景阳冈。天色将晚，酒劲上头，只听一阵狂风，老虎来啦！",
        "幼儿园女教师-温柔甜美": "月亮婆婆升上天空啦，星星宝宝都困啦。小白兔躺在床上，盖好小被子，闭上眼睛。兔妈妈轻轻地唱着摇篮曲：睡吧睡吧，我亲爱的宝贝。",
        "电台主播-平静温柔": "大家好，欢迎收听你的月亮我的心，好男人就是我，我就是：曾小贤。",
        "成熟御姐-温柔暧昧": "小帅哥，今晚有空吗？陪姐姐喝一杯，聊点有意思的。",
        "年轻妈妈-温暖安抚": "从前有座山，山里有座庙，庙里面有个小和尚，小和尚在给老和尚讲故事，他说：从前有座山，山里有座庙，庙里面有个小和尚。",
        "小女孩-尖锐清脆": "一一得一！一二得二！一三得三！我会背乘法口诀啦！老师今天表扬我啦！妈妈说我最棒！",
        "老奶奶-沙哑低沉": "很久很久以前，在山的那边，住着一只会说话的狐狸。它常常在月圆之夜，变成美丽的姑娘，来到村子里。",
        "诗歌朗诵-雄浑有力": "为什么我的眼里常含泪水？因为我对这土地爱得深沉。这土地，这河流，这吹刮着的暴风。",
        "童话风格-甜美夸张": "在一个很冷很冷的夜晚，小女孩擦亮了一根火柴。突然，温暖的火炉出现了！她觉得自己好像坐在火炉旁。",
        "评书风格-抑扬顿挫": "话说那武松，提着哨棒，直奔景阳冈。天色将晚，酒劲上头，只听一阵狂风，老虎来啦！",
        "新闻风格-平静专业": "本台讯，今日凌晨，我国成功发射新一代载人飞船试验船。此次任务验证了多项关键技术，为后续空间站建设奠定基础。",
        "相声风格-夸张幽默": "我这个人啊，最大的优点就是太谦虚。谦虚到什么程度？连谦虚本身都觉得我太谦虚了！",
        "悬疑小说-低沉神秘": "深夜，他独自走在空无一人的小巷。脚步声，回声，还有……另一个人的呼吸声。他猛地回头——什么也没有。",
        "戏剧表演-夸张戏剧": "我疯了！彻底疯了！你们都说我疯了！可疯的是这个世界！清醒的人反而被当成疯子！",
        "法治节目-庄严庄重": "天网恢恢，疏而不漏。任何触犯法律的行为，终将受到公正的审判。正义或许会迟到，但绝不会缺席。",
        "纪录片旁白-低沉磁性": "在这片广袤的非洲草原上，生命与死亡每天都在上演。猎豹的速度，羚羊的敏捷，都是生存的代价。",
        "广告配音-沧桑浑厚": "一杯敬过往，一杯敬远方。传承千年的酿造工艺，只在每一滴醇香。老朋友，值得好酒。",
        "冥想引导师-空灵悠长": "想象你是一片叶子，随风飘落。没有牵挂，没有重量。只有呼吸，只有当下，只有宁静。",
        "ASMR-气声耳语": "现在，让我在你耳边轻声细语。听到我的声音了吗？放松你的头皮，感受每一个毛孔都在呼吸。",
    }

    THEME = gr.themes.Soft(
                primary_hue="orange",
                secondary_hue="cyan",
                neutral_hue="slate",
                )

    CUSTOM_CSS = """
    /* layout */
    #vs-root {max-width: 1180px; margin: 0 auto;}
    #vs-header {padding: 14px 14px 4px 14px;}
    #vs-card {border-radius: 14px; padding: 14px; border: 1px solid rgba(0,0,0,0.08);}

    /* ===== VoiceSculptor palette (from logo) ===== */
    :root, .gradio-container {
    --vs-orange: #FF6A00;
    --vs-orange2:#FFB000;
    --vs-teal:   #00A6C6;
    --vs-blue:   #0B2E8A;
    --vs-teal-a: rgba(0,166,198,.18);
    }

    /* primary button */
    .gr-button-primary, button.primary {
    background: linear-gradient(90deg, var(--vs-orange), var(--vs-orange2)) !important;
    border: none !important;
    color: white !important;
    }
    .gr-button-primary:hover, button.primary:hover {
    filter: brightness(1.03);
    }
    .gr-button-primary:active, button.primary:active {
    filter: brightness(0.98);
    }

    /* links */
    .gradio-container a {
    color: var(--vs-teal) !important;
    }
    .gradio-container a:hover {
    text-decoration: underline;
    }

    /* focus ring / active border for inputs */
    textarea:focus, input:focus {
    border-color: var(--vs-teal) !important;
    box-shadow: 0 0 0 3px var(--vs-teal-a) !important;
    outline: none !important;
    }
    /* some gradio versions wrap inputs in these */
    .gr-input:focus-within, .gr-text-input:focus-within, .gr-box:focus-within {
    border-color: var(--vs-teal) !important;
    box-shadow: 0 0 0 3px var(--vs-teal-a) !important;
    }

    /* accordion highlight */
    .gr-accordion .label, .gr-accordion summary {
    color: var(--vs-blue) !important;
    }
    """

    DEFAULT_STYLE = "评书风格-抑扬顿挫"
    template_choices = [k for k in INSTRUCT_TEMPLATES.keys() if k not in ("default",)]

    BEST_PRACTICE_MD = """
## Best Practice Guide（音色设计）

完整指南请见：Voice Design README  
https://github.com/ASLP-lab/VoiceSculptor/blob/main/docs/voice_design.md  

### 关键约束
- **voice_prompt ≤ 200 字**
- **当前仅支持中文**
- **待合成文本长度 ≥ 5 个字**

### 写法建议
- **具体**：用可感知特质词（低沉/清脆/沙哑/明亮、语速快慢、音量大小等），避免“好听/不错”。  
- **完整**：建议覆盖 **3–4 个维度**（人设/场景 + 性别/年龄 + 音调/语速 + 音质/情绪）。  
- **客观**：描述声音特征与表达方式，避免“我喜欢/很棒”。  
- **不做模仿**：禁止“像某明星/某演员”，只描述声音特质本身。  
- **尽量精炼**：每个词都承载信息，避免重复强调（如“非常非常”）。  

### 参考模板
> - 这是一位男性评书表演者，用传统说唱腔调，以变速节奏和韵律感极强的语速讲述江湖故事，音量时高时低，充满江湖气。  
> - 深夜电台主播，男性、音调偏低、语速偏慢、音量小；情绪平静带点忧伤，语气温柔；音色微哑。
> - 成熟御姐风格，语速正常，音量适中，情绪克制冷静，语气不容置疑的坚定，磁性音色，吐字清晰。


### 细粒度控制提示
- 细粒度控制（年龄/性别/音调/语速/音量/情感等）**建议与指令描述保持一致**，尽量避免相互矛盾（如指令写“低沉慢速”，细粒度却选“音调很高/语速很快”）。
"""

    with gr.Blocks(theme=THEME, css=CUSTOM_CSS) as app:
        with gr.Column(elem_id="vs-root"):
            with gr.Row(elem_id="vs-header"):
                gr.HTML(f"""
                <div style="display:flex; align-items:center; gap:16px;">
                  <img src="{LOGO_URL}"
                       alt="Voice Sculptor Logo"
                       style="width:360px; max-height:130px; object-fit:contain; display:block;" />
                  <div>
                    <div style="font-size:32px; font-weight:700; line-height:1;">Voice Sculptor</div>
                    <div style="opacity:.85; margin-top:6px;">
                      {i18n('An instruct text-to-speech solution based on LLaSA and CosyVoice2 developed by the ASLP lab and collaborators.')}
                    </div>
                  </div>
                </div>
                """)

            with gr.Row():
                # Left: Controls + Guide
                with gr.Column(scale=5, elem_id="vs-card"):
                    gr.Markdown("""
                                ### 🪄 Voice Design（捏音色）
                                <div style="font-size:12px; color:#6B7280; margin-top:6px;">
                                * 音色有随机性；不满意的话可以多生成几次，挑你最喜欢的版本。
                                </div>
                                """)

                    with gr.Accordion("🎭 风格与文本", open=True):
                        instruct_template = gr.Dropdown(
                            choices=template_choices,
                            value=DEFAULT_STYLE,
                            label=i18n("指令风格（必选）"),
                            interactive=True,
                        )

                        instruct_text = gr.Textbox(
                            label=i18n("指令文本"),
                            placeholder=TEXTBOX_PLACEHOLDER,
                            lines=4,
                            value=INSTRUCT_TEMPLATES.get(DEFAULT_STYLE, INSTRUCT_TEMPLATES["default"]),
                        )

                        text = gr.Textbox(
                            label=i18n("待合成文本"),
                            placeholder=TEXTBOX_PLACEHOLDER,
                            lines=4,
                            value=TEXT_REQUIREMENTS.get(DEFAULT_STYLE, TEXT_REQUIREMENTS["default"]),
                        )

                    with gr.Accordion("🎛️ 细粒度声音控制（可选）", open=False):
                        with gr.Row():
                            age_ctrl = gr.Dropdown(label="年龄", choices=["不指定", "小孩", "青年", "中年", "老年"], value="不指定")
                            gender_ctrl = gr.Dropdown(label="性别", choices=["不指定", "男性", "女性"], value="不指定")

                        with gr.Row():
                            pitch_ctrl = gr.Dropdown(
                                label="音调高度",
                                choices=["不指定", "音调很高", "音调较高", "音调中等", "音调较低", "音调很低"],
                                value="不指定",
                            )
                            pitch_var_ctrl = gr.Dropdown(
                                label="音调变化",
                                choices=["不指定", "音调变化很强", "音调变化较强", "音调变化一般", "音调变化较弱", "音调变化很弱"],
                                value="不指定",
                            )

                        with gr.Row():
                            volume_ctrl = gr.Dropdown(
                                label="音量",
                                choices=["不指定", "音量很大", "音量较大", "音量中等", "音量较小", "音量很小"],
                                value="不指定",
                            )
                            speed_ctrl = gr.Dropdown(
                                label="语速",
                                choices=["不指定", "语速很快", "语速较快", "语速中等", "语速较慢", "语速很慢"],
                                value="不指定",
                            )

                        emo_ctrl = gr.Dropdown(
                            label="情感",
                            choices=["不指定", "开心", "生气", "难过", "惊讶", "厌恶", "害怕"],
                            value="不指定",
                        )

                    with gr.Accordion("📚 Best Practice Guide", open=False):
                        gr.Markdown(BEST_PRACTICE_MD)

                    def apply_template(tpl_name):
                        return INSTRUCT_TEMPLATES.get(tpl_name, ""), TEXT_REQUIREMENTS.get(tpl_name, "")

                    instruct_template.change(apply_template, inputs=[instruct_template], outputs=[instruct_text, text])

                # Right: Results + Generate
                with gr.Column(scale=5, elem_id="vs-card"):
                    gr.Markdown("### 🎵 Results")
                    generate = gr.Button("🎧 Generate", variant="primary")
                    audio_output1 = gr.Audio(label=i18n("Generated Audio 1"), type="numpy", interactive=False)
                    audio_output2 = gr.Audio(label=i18n("Generated Audio 2"), type="numpy", interactive=False)
                    audio_output3 = gr.Audio(label=i18n("Generated Audio 3"), type="numpy", interactive=False)

        generate.click(
            fn=inference_select_best3,
            inputs=[text, instruct_text, age_ctrl, gender_ctrl, pitch_ctrl, pitch_var_ctrl, volume_ctrl, speed_ctrl, emo_ctrl],
            outputs=[audio_output1, audio_output2, audio_output3],
        )

    return app


if __name__ == "__main__":
    demo = build_app()
    demo.launch()