Fangjun Kuang

> what do you think about make different logos fot TTS, ASR/STT and speaker identification Yes, that sounds good to me. Would you like to contribute?

> please confirm licence of the original logo The logo is from us and we are publishing all of our work with Apache 2.0 license.

please try https://github.com/k2-fsa/k2/pull/1251

Please refer to the help doc of `k2.intersect_dense`: https://k2-fsa.github.io/k2/python_api/api.html#k2.intersect_dense There are two extra optional arguments: ```python3 def intersect_dense(a_fsas: Fsa, b_fsas: DenseFsaVec, output_beam: float, a_to_b_map: Optional[torch.Tensor] = None, seqframe_idx_name: Optional[str] =...

```python3 arcs = lats.arcs.values()[:, :2] # arcs is a 2-D torch.int32 tensor for idx, (src, dst) in enumerate(arcs.tolist()): # note src is not used and you can replace it with...

> But I'm not sure we can recover the mapping from state to frame by lats.frame only. As I posted above, you can iterate over the arcs; for each arc,...

Note: I have re-edited the demo code.

> Previously, I did not notice the API lats.frame. However, I have recovered this mapping in another way (see code below) Please use a small `lats` to verify that your...

You can note down which state belongs to which frame. For example ``` frame_0_states = [1, 2, 3] frame_1_states = [4, 5, 6, 7, 8, 9] .... ``` You can...

> frame2state: [[0], [1, 2], [13, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15], [16, 18, 17], [19]] `[0]` is the start state and does not...