iddlmZddlmZmZmZddlZddlmZddlm Z gZ e ddededeed eed e d eeeff d Z eGd dZ ddeded e d ee fdZy)) dataclass)ListOptionalTupleN)Tensor)fail_if_no_align log_probstargets input_lengthstarget_lengthsblankreturnc||vrtd|dtj||jdk\r td|T|j d|j d}}tj |f|tj |j}|T|j d|j d}}tj |f|tj |j}|J|Jtjjj|||||\}}||ddtj|jd|dffS) aAlign a CTC label sequence to an emission. .. devices:: CPU CUDA .. properties:: TorchScript Args: log_probs (Tensor): log probability of CTC emission output. Tensor of shape `(B, T, C)`. where `B` is the batch size, `T` is the input length, `C` is the number of characters in alphabet including blank. targets (Tensor): Target sequence. Tensor of shape `(B, L)`, where `L` is the target length. input_lengths (Tensor or None, optional): Lengths of the inputs (max value must each be <= `T`). 1-D Tensor of shape `(B,)`. target_lengths (Tensor or None, optional): Lengths of the targets. 1-D Tensor of shape `(B,)`. blank_id (int, optional): The index of blank symbol in CTC emission. (Default: 0) Returns: Tuple(Tensor, Tensor): Tensor: Label for each time step in the alignment path computed using forced alignment. Tensor: Log probability scores of the labels for each time step. Note: The sequence length of `log_probs` must satisfy: .. math:: L_{\text{log\_probs}} \ge L_{\text{label}} + N_{\text{repeat}} where :math:`N_{\text{repeat}}` is the number of consecutively repeated tokens. For example, in str `"aabbc"`, the number of repeats are `2`. Note: The current version only supports ``batch_size==1``. z4targets Tensor shouldn't contain blank index. Found .z2targets values must be less than the CTC dimensionNr)dtypedevice) ValueErrortorchmaxshapesizefullint64rops torchaudio forced_alignarange) r r r r r batch_sizelengthpathsscoress u/Volumes/fast/ai/experiments/voice-extract-mac/.venv/lib/python3.12/site-packages/torchaudio/functional/_alignment.pyrr s:Z OPWyXYZ[[ yyY__R00MNN&^^A. q0AF  J=& T]TdTde $\\!_gll1oF ZM6U\UcUcd  $$ $  %% %II((55i-YginoME6 &ELLa958CD DDcNeZdZUdZeed< eed< eed< eed< defdZy) TokenSpanz[TokenSpan() Token with time stamps and score. Returned by :py:func:`merge_tokens`. tokenstartendscorerc4|j|jz S)zReturns the time span)r*r))selfs r$__len__zTokenSpan.__len__[sxx$**$$r%N)__name__ __module__ __qualname____doc__int__annotations__floatr.r%r$r'r'Ls5 J J; H9 L&%%r%r'tokensr#c |jdk7s|jdk7r tdt|t|k7r tdtj|tj dg|j tj dg|j }tj|dk7jj}|j}t|dd|ddDcgc]=\}}||x}|k7r.t||||||jj ?}}}|Scc}}w) aRemoves repeated tokens and blank tokens from the given CTC token sequence. Args: tokens (Tensor): Alignment tokens (unbatched) returned from :py:func:`forced_align`. Shape: `(time, )`. scores (Tensor): Alignment scores (unbatched) returned from :py:func:`forced_align`. Shape: `(time, )`. When computing the token-size score, the given score is averaged across the corresponding time span. Returns: list of TokenSpan Example: >>> aligned_tokens, scores = forced_align(emission, targets, input_lengths, target_lengths) >>> token_spans = merge_tokens(aligned_tokens[0], scores[0]) rz(`tokens` and `scores` must be 1D Tensor.z.`tokens` and `scores` must be the same length.r)r)prependappendrN)r(r)r*r+)ndimrlenrdifftensorrnonzerosqueezetolistzipr'meanitem) r7r#r r=changes_wo_blankr)r*r(spanss r$ merge_tokensrG`s/"{{a6;;!+CDD 6{c&k!IJJ :: bT&--@WYVZcicpcpIq D}}dai199;BBD ]]_F.s35Eab5IJ JJE3E] "Eu , U6%;L;Q;Q;S;X;X;Z[J   L  s=AE)NNr)r) dataclassesrtypingrrrrrtorchaudio._extensionr__all__r3rr'rGr6r%r$rLs!(( 2 '+'+ =E=E =EF#=EV$ =E  =E  66> =E=E@ %% %&    DO r%