paddlespeech.s2t.utils.ctc_utils module

paddlespeech.s2t.utils.ctc_utils.forced_align(ctc_probs: Tensor, y: Tensor, blank_id=0) → List[int]

ctc forced alignment.

https://distill.pub/2017/ctc/

Args:

ctc_probs (paddle.Tensor): hidden state sequence, 2d tensor (T, D) y (paddle.Tensor): label id sequence tensor, 1d tensor (L) blank_id (int): blank symbol index

Returns:

List[int]: best alignment result, (T).

paddlespeech.s2t.utils.ctc_utils.insert_blank(label: ndarray, blank_id: int = 0) → ndarray

Insert blank token between every two label token.

"abcdefg" -> "-a-b-c-d-e-f-g-"

Args:

label ([np.ndarray]): label ids, List[int], (L). blank_id (int, optional): blank id. Defaults to 0.

Returns:

[np.ndarray]: (2L+1).

paddlespeech.s2t.utils.ctc_utils.remove_duplicates_and_blank(hyp: List[int], blank_id=0) → List[int]

ctc alignment to ctc label ids.

"abaa-acee-" -> "abaace"

Args:

hyp (List[int]): hypotheses ids, (L) blank_id (int, optional): blank id. Defaults to 0.

Returns:

List[int]: remove dupicate ids, then remove blank id.