I don't intend to support ONNX for the LSTM since I don't see a path forward with ONNX to supporting models with non-trivial receptive fields like the WaveNet model. If ONNX show interest in supporting real-time audio applications, I may take it back, but there's enough that's different about implementing for RTA that I don't see them picking it up--most ML compute is optimized for large-batch, high-latency applications.
I don't intend to support ONNX for the LSTM since I don't see a path forward with ONNX to supporting models with non-trivial receptive fields like the WaveNet model. If ONNX show interest in supporting real-time audio applications, I may take it back, but there's enough that's different about implementing for RTA that I don't see them picking it up--most ML compute is optimized for large-batch, high-latency applications.