SmileSky

您能给一个测试脚本吗，这样统计的性能数据可以尽可能和你那边一致。 我这边没有gpu设备，cpu是i5的也比较老的，实时推理有点慢，但是可接受范围，合成“你好呀，我们是一群追逐梦想的人。”15个字大概需要6秒，输出的音频只有两三秒，显然达不到实时性。

不知道模型是否还可以继续压缩。

我使用的是主分支推理，我的cpu是Intel(R) Core(TM) i5-4670T CPU @ 2.30GHz，4核心。

```python import time def test_list_operations(): lst = [i for i in range(100000)] start_time = time.time() for _ in range(1000): lst.append(len(lst)) lst.remove(lst[0]) end_time = time.time() print("List operations took", end_time - start_time,...

Welcome to follow the Screen Reader implemented by me: https://github.com/mzdk100/rigela.git

The functions are already quite comprehensive, welcome to participate and contribute.

Hope to support Chinese.

If using the tokio library and enabling the "process" feature, there will also be a STATUS_ENTRYPOINT_NOT_FOUND. cargo.toml： ```toml [dependencies] native-windows-derive = "1.0.5" native-windows-gui = "1.0.13" tokio = { version =...

Thank you for your reply. Training with Candle is a bit difficult for me. I may choose to train YOLOv8 on PyTorch, then convert the checkpoint of PT to safetensors...

I haven't started trying yet.