Update app.py

app.py

@@ -1,82 +1,64 @@
-import os
-import gradio as gr
-import numpy as np
-import soundfile as sf
-import tempfile
-
-from huggingface_hub import snapshot_download
-import onnxruntime as ort
-
-
-# =========================
-# 1. 自动下载模型（关键）
-# =========================
-MODEL_DIR = snapshot_download(
-    repo_id="OpenMOSS-Team/MOSS-TTS-Nano-100M-ONNX",
-    local_dir="./models",
-    local_dir_use_symlinks=False
-)
-
-
-# =========================
-# 2. ONNX TTS 封装（简化可运行结构）
-# =========================
-class MOSSTTS:
-    def __init__(self, model_dir):
-        self.prefill = ort.InferenceSession(
-            f"{model_dir}/moss_tts_prefill.onnx",
-            providers=["CPUExecutionProvider"]
-        )
-
-        self.decode = ort.InferenceSession(
-            f"{model_dir}/moss_tts_decode_step.onnx",
-            providers=["CPUExecutionProvider"]
-        )
-
-    def infer(self, text):
-        """
-        ⚠️ 注意：这里是最小可跑demo结构
-        实际项目需要 tokenizer + codec
-        """
-
-        # fake token（占位）
-        input_ids = np.array([[1, 2, 3]], dtype=np.int64)
-
-        self.prefill.run(None, {"input_ids": input_ids})
-
-        # fake audio
-        wav = np.random.randn(16000 * 3).astype(np.float32)
-        sr = 16000
+import sys, os, time, io, wave
+sys.path.insert(0, "/app")
+os.environ["OMP_NUM_THREADS"] = "2"
 
-        return wav, sr
-
-
-# =========================
-# 3. 初始化模型
-# =========================
-tts = MOSSTTS(MODEL_DIR)
-
-
-# =========================
-# 4. 推理函数
-# =========================
-def generate(text):
-    wav, sr = tts.infer(text)
-
-    out_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False).name
-    sf.write(out_file, wav, sr)
-
-    return out_file
-
-
-# =========================
-# 5. Gradio UI
-# =========================
-demo = gr.Interface(
-    fn=generate,
-    inputs=gr.Textbox(label="Text"),
-    outputs=gr.Audio(label="Output Audio"),
-    title="MOSS-TTS-Nano ONNX (CPU)",
-)
-
-demo.launch(server_name="0.0.0.0", server_port=7860)
+import numpy as np
+import gradio as gr
+from onnx_tts_runtime import OnnxTtsRuntime, _merge_audio_channels, _concat_waveforms, _write_waveform_to_wav
+
+MODEL_DIR = "/app/models"
+BUILTIN_VOICES = None
+runtime = None
+
+def load_runtime():
+    global runtime, BUILTIN_VOICES
+    if runtime is not None:
+        return runtime
+    runtime = OnnxTtsRuntime(
+        model_dir=MODEL_DIR,
+        thread_count=2,
+        max_new_frames=375,
+        execution_provider="cpu",
+    )
+    BUILTIN_VOICES = [v["voice"] for v in runtime.list_builtin_voices()]
+    return runtime
+
+def synthesize(text, voice, audio_path, sample_mode, max_frames):
+    rt = load_runtime()
+    t0 = time.time()
+    result = rt.synthesize(
+        text=text,
+        voice=voice if not audio_path else None,
+        prompt_audio_path=audio_path if audio_path else None,
+        sample_mode=sample_mode,
+        do_sample=(sample_mode != "greedy"),
+        streaming=True,
+        max_new_frames=int(max_frames),
+        enable_wetext=False,
+        enable_normalize_tts_text=False,
+    )
+    elapsed = time.time() - t0
+    sr = result["sample_rate"]
+    wav_path = result["audio_path"]
+    return wav_path, f"Done in {elapsed:.1f}s | {sr}Hz | {int(result['audio_token_ids'].shape[0])} frames"
+
+with gr.Blocks(title="MOSS-TTS-Nano ONNX") as demo:
+    gr.Markdown("# MOSS-TTS-Nano-100M-ONNX\nCPU-only TTS with voice cloning. First run downloads ~730MB model.")
+    with gr.Row():
+        with gr.Column():
+            text_in = gr.Textbox(label="Text", value="Hello, welcome to MOSS TTS Nano.", lines=3)
+            with gr.Row():
+                voice_in = gr.Dropdown(choices=["Junhao","Zhiming","Weiguo","Xiaoyu","Yuewen","Lingyu","Trump","Ava","Bella","Adam","Nathan","Soyo","Saki","Mortis","Umiri","Mei","Anon","Arisa"], value="Junhao", label="Voice (overridden by ref audio)")
+                ref_audio = gr.Audio(label="Reference Audio (optional, for voice cloning)", type="filepath")
+            with gr.Row():
+                sample_mode = gr.Dropdown(choices=["fixed","greedy","full"], value="fixed", label="Sample Mode")
+                max_frames = gr.Slider(16, 750, value=375, step=1, label="Max Frames")
+            btn = gr.Button("Synthesize", variant="primary")
+        with gr.Column():
+            audio_out = gr.Audio(label="Generated Audio", type="filepath")
+            info_out = gr.Textbox(label="Info")
+    btn.click(fn=synthesize, inputs=[text_in, voice_in, ref_audio, sample_mode, max_frames], outputs=[audio_out, info_out])
+
+if __name__ == "__main__":
+    load_runtime()
+    demo.launch()