For kaldi matrix, only float matrix are supported. Please pass "tf.float32" in through the element-type argument.
For compressed kaldi matrix, only compression method 2 (kSpeechFeature) is supported.
For alignment, user need to specify weather is reading posteriors or pdfs by the argument is_reading_post. Operator decode_kaldi_ali produces alignment pdfs output, in format of a one dimension int32 tensor (a int32 tensor). Please pass "tf.int32" in through the element-type argument.
Contributions are welcome. Feel free to fork and send pull request or to create issues.
There are some examples under the example directory. To run them, please modify the library path (in contents of the python files) to the correct path.