update

src/voice/cosyvoice-opt-test.py

@@ -0,0 +1,66 @@
+#!/usr/bin/env python3
+"""CosyVoice2 性能对比：非流式 / 流式 / 流式+jit
+
+用法: py -3 src/voice/cosyvoice-opt-test.py [--jit] [--stream]
+"""
+import os
+import sys
+import time
+import subprocess
+
+BASE = os.path.dirname(os.path.abspath(__file__))
+MODEL_DIR = os.path.join(BASE, "models", "CosyVoice2-0.5B")
+REF_WAV = os.path.join(BASE, "models", "melo-tts", "ref_female.wav")
+REF_TEXT = "你好，我是小雅，今天想和你聊聊天。"
+
+DO_JIT = "--jit" in sys.argv
+DO_STREAM = "--stream" in sys.argv
+
+sys.path.insert(0, os.path.join(BASE, "cosyvoice-src"))
+sys.path.insert(0, os.path.join(BASE, "cosyvoice-src", "third_party", "Matcha-TTS"))
+
+def gpu_mem():
+    try:
+        out = subprocess.run(
+            ["nvidia-smi", "--query-gpu=memory.used,memory.total", "--format=csv,noheader,nounits"],
+            capture_output=True, text=True, timeout=5,
+        ).stdout.strip()
+        return out
+    except Exception:
+        return "N/A"
+
+def main():
+    tag = "stream" if DO_STREAM else "batch"
+    if DO_JIT:
+        tag += "+jit"
+    print(f"[opt] 配置: {tag} | GPU: {gpu_mem()} MiB", flush=True)
+    t0 = time.time()
+
+    from cosyvoice.cli.cosyvoice import AutoModel
+    cosyvoice = AutoModel(model_dir=MODEL_DIR, fp16=True, load_jit=DO_JIT)
+    print(f"[opt] 模型加载: {time.time()-t0:.1f}s | GPU: {gpu_mem()} MiB", flush=True)
+
+    text = "你好，我是小雅，很高兴认识你。今天我们聊聊白龙马的奇妙旅程吧。"
+
+    import torch
+    t1 = time.time()
+    first_chunk_t = None
+    chunks = []
+    gen = cosyvoice.inference_zero_shot(text, REF_TEXT, REF_WAV, stream=DO_STREAM)
+    for i, j in enumerate(gen):
+        if i == 0:
+            first_chunk_t = time.time() - t1
+            print(f"[opt] 首包: {first_chunk_t:.1f}s | GPU: {gpu_mem()} MiB", flush=True)
+        chunks.append(j['tts_speech'])
+    elapsed = time.time() - t1
+    wav = torch.cat(chunks, dim=1)
+    duration = wav.shape[1] / cosyvoice.sample_rate
+    print(f"[opt] 合成: {elapsed:.1f}s | 音频 {duration:.1f}s | RTF {elapsed/duration:.2f}"
+          f" | 首包 {first_chunk_t:.1f}s | GPU: {gpu_mem()} MiB", flush=True)
+
+    import soundfile as sf
+    sf.write(f"cosyvoice_{tag}.wav", wav.squeeze(0).cpu().numpy(), cosyvoice.sample_rate)
+    print(f"[opt] 已保存: cosyvoice_{tag}.wav", flush=True)
+
+if __name__ == "__main__":
+    main()