happyme531/F5-TTS-RKNN2

F5-TTS-RKNN2

(English README see below)

在RK3588上运行超高质量的F5-TTS文字转语音/零样本音色克隆模型!

• 推理速度(RK3588, 生成9秒音频): 每次迭代用时11s, 迭代32步, 总用时352s
• 内存占用(RK3588): 2.2GB

使用方法

1. 克隆或者下载此仓库到本地. 模型较大, 请确保有足够的磁盘空间.

2. 安装依赖

pip install "numpy<2" rknn-toolkit-lite2 jieba torch onnxruntime soundfile pydub pypinyin tqdm

4. 运行

python F5-TTS-ONNX-Inference-rknn2.py

你可以修改F5-TTS-ONNX-Inference-rknn2.py中的文本等参数来生成不同的音频。

模型转换

1. 下载ONNX模型文件

2. 安装依赖

pip install "numpy<2" rknn-toolkit2==2.3.0 onnx onnxruntime

3. 转换模型

python convert_opset.py
python convert_F5_Transformer_opset19.py

已知问题

• 由于RKNN2不支持动态输入，这里把序列长度固定为了1536，并通过缩放音频速度来补偿。在差距不大的情况下效果可以接受。
• 模型中DiT中的RoPE位置编码部分有一个Transpose操作无法在NPU上运行，造成推理速度下降~15%。这个问题应该可以通过修改原模型来解决，但我懒得改了，因为改完之后推理还是会非常慢，因为序列长度实在太长了。
• 只有DiT部分使用了NPU，其他部分都是CPU推理，但其他部分运行速度快，总体上不会对推理速度有太大影响。

参考

• F5-TTS-ONNX
• F5-TTS

English README

Run the ultra-high-quality F5-TTS text-to-speech / zero-shot voice cloning model on RK3588!

• Inference Speed (on RK3588, generating 9 seconds of audio): 11s per iteration, 32 iterations, total time ~352s
• Memory Usage (on RK3588): 2.2GB

Usage

1. Clone or download this repository locally. The models are large, ensure you have sufficient disk space.

2. Install dependencies:

   pip install "numpy<2" rknn-toolkit-lite2 jieba torch onnxruntime soundfile pydub pypinyin tqdm
   

3. Run:

   python F5-TTS-ONNX-Inference-rknn2.py
   

   You can modify parameters such as the text within F5-TTS-ONNX-Inference-rknn2.py to generate different audio.

Model Conversion

1. Download the ONNX model files.

2. Install dependencies:

   pip install "numpy<2" rknn-toolkit2==2.3.0 onnx onnxruntime
   

3. Convert the models:

   python convert_opset.py
   python convert_F5_Transformer_opset19.py
   

Known Issues

• Due to RKNN2 limitations with dynamic inputs, the sequence length is fixed at 1536. Audio speed scaling is used to compensate. The effect is acceptable when the difference isn't significant.
• A Transpose operation within the RoPE (Rotary Positional Embedding) part of the DiT (Diffusion Transformer) model cannot run on the NPU, causing an approximate 15% decrease in inference speed. This could potentially be resolved by modifying the original model, but I chose not to, as inference would still be very slow due to the extremely long sequence length.
• Only the DiT part utilizes the NPU; other parts run on the CPU. However, these CPU parts are fast and do not significantly impact the overall inference speed.

References

• F5-TTS-ONNX
• F5-TTS