Add TODO task lists.

README.md

@@ -9,3 +9,15 @@ Wrap kaldi's feature computations to Python with PyTorch support.
 ```bash
 pip install kaldifeat
 ```
+
+# TODOs
+
+- [ ] Add Python interface
+- [ ] Support torch.device so that it can switch between CUDA and CPU
+- [ ] Add unit tests
+- [ ] Set up GitHub actions
+- [ ] Benchmark its speed and compare it with Kaldi
+- [ ] Support batch processing of multiple waves
+- [ ] Handle non-default parameters
+- [ ] Support MFCC and other features available in Kaldi
+- [ ] Publish it to PyPI

kaldifeat/csrc/feature-window.cc

@@ -134,7 +134,12 @@ torch::Tensor Dither(const torch::Tensor &wave, float dither_value) {
   if (dither_value == 0.0f) wave;
 
   torch::Tensor rand_gauss = torch::randn_like(wave);
+#if 1
   return wave + rand_gauss * dither_value;
+#else
+  // use in-place version of wave and change its to pointer type
+  wave_->add_(rand_gauss, dither_value);
+#endif
 }
 
 void Preemphasize(float preemph_coeff, torch::Tensor *wave) {