shixian
2024-12-05 026b8e3fdc981ceeac18257319fb4b1b7db2f8b5
update sensevoice small with timestamp
2个文件已修改
1个文件已添加
156 ■■■■■ 已修改文件
examples/industrial_data_pretraining/sense_voice/demo.py 30 ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史
funasr/models/sense_voice/model.py 61 ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史
funasr/models/sense_voice/utils/ctc_alignment.py 65 ●●●●● 补丁 | 查看 | 原始文档 | blame | 历史
examples/industrial_data_pretraining/sense_voice/demo.py
@@ -29,6 +29,21 @@
text = rich_transcription_postprocess(res[0]["text"])
print(text)
# en with timestamp
res = model.generate(
    input=f"{model.model_path}/example/en.mp3",
    cache={},
    language="auto",  # "zn", "en", "yue", "ja", "ko", "nospeech"
    use_itn=True,
    batch_size_s=60,
    merge_vad=True,  #
    merge_length_s=15,
    output_timestamp=True,
)
print(res)
text = rich_transcription_postprocess(res[0]["text"])
print(text)
# zh
res = model.generate(
    input=f"{model.model_path}/example/zh.mp3",
@@ -42,6 +57,21 @@
text = rich_transcription_postprocess(res[0]["text"])
print(text)
# zh with timestamp
res = model.generate(
    input=f"{model.model_path}/example/zh.mp3",
    cache={},
    language="auto",  # "zn", "en", "yue", "ja", "ko", "nospeech"
    use_itn=True,
    batch_size_s=60,
    merge_vad=True,  #
    merge_length_s=15,
    output_timestamp=True,
)
print(res)
text = rich_transcription_postprocess(res[0]["text"])
print(text)
# yue
res = model.generate(
    input=f"{model.model_path}/example/yue.mp3",
funasr/models/sense_voice/model.py
@@ -19,6 +19,7 @@
from funasr.models.paraformer.search import Hypothesis
from funasr.models.sense_voice.utils.ctc_alignment import ctc_forced_align
class SinusoidalPositionEncoder(torch.nn.Module):
@@ -857,6 +858,8 @@
        use_itn = kwargs.get("use_itn", False)
        textnorm = kwargs.get("text_norm", None)
        output_timestamp = kwargs.get("output_timestamp", False)
        if textnorm is None:
            textnorm = "withitn" if use_itn else "woitn"
        textnorm_query = self.embed(
@@ -905,18 +908,70 @@
            # Change integer-ids to tokens
            text = tokenizer.decode(token_int)
            result_i = {"key": key[i], "text": text}
            results.append(result_i)
            # result_i = {"key": key[i], "text": text}
            # results.append(result_i)
            if ibest_writer is not None:
                ibest_writer["text"][key[i]] = text
            if output_timestamp:
                from itertools import groupby
                timestamp = []
                tokens = tokenizer.text2tokens(text)[4:]
                logits_speech = self.ctc.softmax(encoder_out)[i, 4:encoder_out_lens[i].item(), :]
                pred = logits_speech.argmax(-1).cpu()
                logits_speech[pred==self.blank_id, self.blank_id] = 0
                align = ctc_forced_align(
                    logits_speech.unsqueeze(0).float(),
                    torch.Tensor(token_int[4:]).unsqueeze(0).long().to(logits_speech.device),
                    (encoder_out_lens-4).long(),
                    torch.tensor(len(token_int)-4).unsqueeze(0).long().to(logits_speech.device),
                    ignore_id=self.ignore_id,
                )
                pred = groupby(align[0, :encoder_out_lens[0]])
                _start = 0
                token_id = 0
                ts_max = encoder_out_lens[i] - 4
                for pred_token, pred_frame in pred:
                    _end = _start + len(list(pred_frame))
                    if pred_token != 0:
                        ts_left = max((_start*60-30)/1000, 0)
                        ts_right = min((_end*60-30)/1000, (ts_max*60-30)/1000)
                        timestamp.append([tokens[token_id], ts_left, ts_right])
                        token_id += 1
                    _start = _end
                timestamp = self.post(timestamp)
                result_i = {"key": key[i], "text": text, "timestamp": timestamp}
                results.append(result_i)
            else:
                result_i = {"key": key[i], "text": text}
                results.append(result_i)
        return results, meta_data
    def post(self, timestamp):
        timestamp_new = []
        for i, t in enumerate(timestamp):
            word, start, end = t
            if word == '▁':
                continue
            if i == 0:
                # timestamp_new.append([word, start, end])
                timestamp_new.append([int(start*1000), int(end*1000)])
            elif word.startswith("▁") or len(word) == 1 or not word[1].isalpha():
                word = word[1:]
                # timestamp_new.append([word, start, end])
                timestamp_new.append([int(start*1000), int(end*1000)])
            else:
                # timestamp_new[-1][0] += word
                timestamp_new[-1][1] = int(end*1000)
        return timestamp_new
    def export(self, **kwargs):
        from .export_meta import export_rebuild_model
        from export_meta import export_rebuild_model
        if "max_seq_len" not in kwargs:
            kwargs["max_seq_len"] = 512
        models = export_rebuild_model(model=self, **kwargs)
        return models
        return results, meta_data
funasr/models/sense_voice/utils/ctc_alignment.py
New file
@@ -0,0 +1,65 @@
import torch
def ctc_forced_align(
    log_probs: torch.Tensor,
    targets: torch.Tensor,
    input_lengths: torch.Tensor,
    target_lengths: torch.Tensor,
    blank: int = 0,
    ignore_id: int = -1,
) -> torch.Tensor:
    """Align a CTC label sequence to an emission.
    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):
            Lengths of the inputs (max value must each be <= `T`). 1-D Tensor of shape `(B,)`.
        target_lengths (Tensor):
            Lengths of the targets. 1-D Tensor of shape `(B,)`.
        blank_id (int, optional): The index of blank symbol in CTC emission. (Default: 0)
        ignore_id (int, optional): The index of ignore symbol in CTC emission. (Default: -1)
    """
    targets[targets == ignore_id] = blank
    batch_size, input_time_size, _ = log_probs.size()
    bsz_indices = torch.arange(batch_size, device=input_lengths.device)
    _t_a_r_g_e_t_s_ = torch.cat(
        (
            torch.stack((torch.full_like(targets, blank), targets), dim=-1).flatten(start_dim=1),
            torch.full_like(targets[:, :1], blank),
        ),
        dim=-1,
    )
    diff_labels = torch.cat(
        (
            torch.as_tensor([[False, False]], device=targets.device).expand(batch_size, -1),
            _t_a_r_g_e_t_s_[:, 2:] != _t_a_r_g_e_t_s_[:, :-2],
        ),
        dim=1,
    )
    neg_inf = torch.tensor(float("-inf"), device=log_probs.device, dtype=log_probs.dtype)
    padding_num = 2
    padded_t = padding_num + _t_a_r_g_e_t_s_.size(-1)
    best_score = torch.full((batch_size, padded_t), neg_inf, device=log_probs.device, dtype=log_probs.dtype)
    best_score[:, padding_num + 0] = log_probs[:, 0, blank]
    best_score[:, padding_num + 1] = log_probs[bsz_indices, 0, _t_a_r_g_e_t_s_[:, 1]]
    backpointers = torch.zeros((batch_size, input_time_size, padded_t), device=log_probs.device, dtype=targets.dtype)
    for t in range(1, input_time_size):
        prev = torch.stack(
            (best_score[:, 2:], best_score[:, 1:-1], torch.where(diff_labels, best_score[:, :-2], neg_inf))
        )
        prev_max_value, prev_max_idx = prev.max(dim=0)
        best_score[:, padding_num:] = log_probs[:, t].gather(-1, _t_a_r_g_e_t_s_) + prev_max_value
        backpointers[:, t, padding_num:] = prev_max_idx
    l1l2 = best_score.gather(
        -1, torch.stack((padding_num + target_lengths * 2 - 1, padding_num + target_lengths * 2), dim=-1)
    )
    path = torch.zeros((batch_size, input_time_size), device=best_score.device, dtype=torch.long)
    path[bsz_indices, input_lengths - 1] = padding_num + target_lengths * 2 - 1 + l1l2.argmax(dim=-1)
    for t in range(input_time_size - 1, 0, -1):
        target_indices = path[:, t]
        prev_max_idx = backpointers[bsz_indices, t, target_indices]
        path[:, t - 1] += target_indices - prev_max_idx
    alignments = _t_a_r_g_e_t_s_.gather(dim=-1, index=(path - padding_num).clamp(min=0))
    return alignments