app.py

import os, time, urllib.request
os.environ["OMP_NUM_THREADS"] = "1"

BASE = "https://raw.githubusercontent.com/OpenMOSS/MOSS-TTS-Nano/main"
for fn in ["onnx_tts_runtime.py", "ort_cpu_runtime.py",
            "text_normalization_pipeline.py", "tts_robust_normalizer_single_script.py"]:
    if not os.path.isfile(fn):
        urllib.request.urlretrieve(f"{BASE}/{fn}", fn)

os.makedirs("moss_tts_nano", exist_ok=True)
for sub, content in [
    ("__init__.py", '__version__ = "0.1.0"\n'),
    ("defaults.py", "from pathlib import Path\nREPO_ROOT = Path(__file__).resolve().parents[1]\nDEFAULT_OUTPUT_DIR = REPO_ROOT / 'generated_audio'\n"),
]:
    if not os.path.isfile(f"moss_tts_nano/{sub}"):
        with open(f"moss_tts_nano/{sub}", "w") as f:
            f.write(content)

import gradio as gr
from onnx_tts_runtime import OnnxTtsRuntime

RT = None
VOICES = ["Junhao","Zhiming","Weiguo","Xiaoyu","Yuewen","Lingyu","Trump","Ava","Bella","Adam","Nathan","Soyo","Saki","Mortis","Umiri","Mei","Anon","Arisa"]

def get_rt():
    global RT
    if RT is None:
        RT = OnnxTtsRuntime(os.environ.get("MOSS_MODEL_DIR", "models"), thread_count=1, execution_provider="cpu")
    return RT

def synth(text, voice, ref, smode, mf, seed, at, ak, ap, ar):
    rt = get_rt()
    gd = rt.manifest["generation_defaults"]
    gd.update(max_new_frames=int(mf), sample_mode=smode, do_sample=smode!="greedy",
              audio_temperature=float(at), audio_top_k=int(ak), audio_top_p=float(ap),
              audio_repetition_penalty=float(ar))
    if seed >= 0:
        import numpy as np
        rt.rng = np.random.default_rng(int(seed))
    t0 = time.time()
    r = rt.synthesize(text=text, voice=voice if not ref else None,
                      prompt_audio_path=ref if ref else None,
                      sample_mode=smode, streaming=True, max_new_frames=int(mf),
                      enable_wetext=False, enable_normalize_tts_text=False)
    return r["audio_path"], f"完成 — {time.time()-t0:.1f}秒 | {r['sample_rate']}Hz"

with gr.Blocks(title="MOSS-TTS-Nano ONNX", theme=gr.themes.Soft()) as demo:
    gr.Markdown("## 🗣️ MOSS-TTS-Nano-100M-ONNX\n纯 CPU 离线 TTS，支持语音克隆。首次运行自动下载 ~730MB 模型。")
    with gr.Row():
        with gr.Column(scale=1):
            text = gr.Textbox(label="合成文本", value="你好，欢迎使用 MOSS 语音合成。", lines=4)
            with gr.Row():
                voice = gr.Dropdown(choices=VOICES, value="Junhao", label="内置声音（上传参考音频后切换为语音克隆）")
                ref = gr.Audio(label="参考音频（可选）", type="filepath")
            with gr.Row():
                smode = gr.Dropdown(choices=["greedy — 贪心解码","fixed — 内置参数（推荐）","full — 自定义采样"], value="fixed — 内置参数（推荐）", label="采样模式")
                frames = gr.Slider(16, 750, value=375, step=1, label="最大帧数")
            seed = gr.Slider(-1, 9999, value=-1, step=1, label="随机种子（-1=随机）")
            btn = gr.Button("🎤 生成语音", variant="primary", size="lg")
            with gr.Accordion("⚙️ 高级采样参数（full 模式生效）", open=False):
                at = gr.Slider(0.1, 2.0, 0.8, 0.05, label="温度 audio_temperature")
                ak = gr.Slider(1, 100, 25, 1, label="Top-K audio_top_k")
                ap = gr.Slider(0.1, 1.0, 0.95, 0.05, label="Top-P audio_top_p")
                ar = gr.Slider(1.0, 2.0, 1.2, 0.05, label="重复惩罚 audio_repetition_penalty")
        with gr.Column(scale=1):
            audio_out = gr.Audio(label="生成结果", type="filepath")
            info = gr.Textbox(label="状态")
    btn.click(fn=lambda t,v,r,m,f,s,at,ak,ap,ar: synth(t,v,r,m.split(" — ")[0],f,s,at,ak,ap,ar),
              inputs=[text,voice,ref,smode,frames,seed,at,ak,ap,ar],
              outputs=[audio_out,info])

if __name__ == "__main__":
    get_rt()
    demo.launch(server_name="0.0.0.0",mcp_server=True)