Hello, I have downloaded the DeepFilterNet3_onnx.tar.gz model from the official website. It contains three ONNX models. I want to use Python for pre-processing and post-processing to obtain the final audio.
I referenced some official Rust code and other people's inference scripts, including the open-source source code of deepfilternet-serverless on GitHub. While I obtained results from all of them, the noise reduction was very poor. Only SpeechDenoiser's single ONNX inference performance matched the official results, and the effect was also good. However, I couldn't figure out how it generated the single ONNX. The author's ONNX file crashed when using ONNXDirectML and converting to MNN, likely due to the use of a fusion operator. The official model didn't have this problem. I'm hoping to find a correct way to use the official model. Are there any related pre-processing and post-processing codes I can refer to? Or have others successfully used the official ONNX for inference, and whether the results were consistent with the official noise reduction?
Hello, I have downloaded the DeepFilterNet3_onnx.tar.gz model from the official website. It contains three ONNX models. I want to use Python for pre-processing and post-processing to obtain the final audio.
I referenced some official Rust code and other people's inference scripts, including the open-source source code of deepfilternet-serverless on GitHub. While I obtained results from all of them, the noise reduction was very poor. Only SpeechDenoiser's single ONNX inference performance matched the official results, and the effect was also good. However, I couldn't figure out how it generated the single ONNX. The author's ONNX file crashed when using ONNXDirectML and converting to MNN, likely due to the use of a fusion operator. The official model didn't have this problem. I'm hoping to find a correct way to use the official model. Are there any related pre-processing and post-processing codes I can refer to? Or have others successfully used the official ONNX for inference, and whether the results were consistent with the official noise reduction?