aishell example

游雁

2024-02-19 94de39dde2e616a01683c518023d0fab72b4e103

 funasr/utils/timestamp_tools.py

@@ -1,137 +1,161 @@
import torch
import copy
import codecs
import logging
import argparse
import numpy as np
from typing import Any, List, Tuple, Union
# import edit_distance
from itertools import zip_longest

def cut_interval(alphas: torch.Tensor, start: int, end: int, tail: bool):
    if not tail:
        if end == start + 1:
            cut = (end + start) / 2.0
        else:
            alpha = alphas[start+1: end].tolist()
            reverse_steps = 1
            for reverse_alpha in alpha[::-1]:
                if reverse_alpha > 0.35:
                    reverse_steps += 1
                else:
                    break
            cut = end - reverse_steps
    else:
        if end != len(alphas) - 1:
            cut = end + 1
        else:
            cut = start + 1
    return float(cut)

def time_stamp_lfr6(alphas: torch.Tensor, speech_lengths: torch.Tensor, raw_text: List[str], begin: int = 0, end: int = None):
    time_stamp_list = []
    alphas = alphas[0]
    text = copy.deepcopy(raw_text)
    if end is None:
        time = speech_lengths * 60 / 1000
        sacle_rate = (time / speech_lengths[0]).tolist()
    else:
        time = (end - begin) / 1000
        sacle_rate = (time / speech_lengths[0]).tolist()
def cif_wo_hidden(alphas, threshold):
    batch_size, len_time = alphas.size()
    # loop varss
    integrate = torch.zeros([batch_size], device=alphas.device)
    # intermediate vars along time
    list_fires = []
    for t in range(len_time):
        alpha = alphas[:, t]
        integrate += alpha
        list_fires.append(integrate)
        fire_place = integrate >= threshold
        integrate = torch.where(fire_place,
                                integrate - torch.ones([batch_size], device=alphas.device)*threshold,
                                integrate)
    fires = torch.stack(list_fires, 1)
    return fires

    predictor = (alphas > 0.5).int()
    fire_places = torch.nonzero(predictor == 1).squeeze(1).tolist()
    
    cuts = []
    npeak = int(predictor.sum())
    nchar = len(raw_text)
    if npeak - 1 == nchar:
        fire_places = torch.where((alphas > 0.5) == 1)[0].tolist()
        for i in range(len(fire_places)):
            if fire_places[i] < len(alphas) - 1:
                if 0.05 < alphas[fire_places[i]+1] < 0.5:
                    fire_places[i] += 1
    elif npeak < nchar:
        lost_num = nchar - npeak
        lost_fire = speech_lengths[0].tolist() - fire_places[-1]
        interval_distance = lost_fire // (lost_num + 1)
        for i in range(1, lost_num + 1):
            fire_places.append(fire_places[-1] + interval_distance)
    elif npeak - 1 > nchar:
        redundance_num = npeak - 1 - nchar
        for i in range(redundance_num):
            fire_places.pop() 

    cuts.append(0)
    start_sil = True
    if start_sil:
        text.insert(0, '<sil>')

    for i in range(len(fire_places)-1):
        cuts.append(cut_interval(alphas, fire_places[i], fire_places[i+1], tail=(i==len(fire_places)-2)))

    for i in range(2, len(fire_places)-2):
        if fire_places[i-2] == fire_places[i-1] - 1 and fire_places[i-1] != fire_places[i] - 1:
            cuts[i-1] += 1

    if cuts[-1] != len(alphas) - 1:
        text.append('<sil>')
        cuts.append(speech_lengths[0].tolist())
    cuts.insert(-1, (cuts[-1] + cuts[-2]) * 0.5)
    sec_fire_places = np.array(cuts) * sacle_rate
    for i in range(1, len(sec_fire_places) - 1):
        start, end = sec_fire_places[i], sec_fire_places[i+1]
        if i == len(sec_fire_places) - 2:
            end = time
        time_stamp_list.append([int(round(start, 2) * 1000) + begin, int(round(end, 2) * 1000) + begin])
        text = text[1:]
    if npeak - 1 == nchar or npeak > nchar:
        return time_stamp_list[:-1]
    else:
        return time_stamp_list

def time_stamp_lfr6_pl(us_alphas, us_cif_peak, char_list, begin_time=0.0, end_time=None):
def ts_prediction_lfr6_standard(us_alphas, 
                       us_peaks, 
                       char_list, 
                       vad_offset=0.0, 
                       force_time_shift=-1.5,
                       sil_in_str=True
                       ):
    if not len(char_list):
        return "", []
    START_END_THRESHOLD = 5
    MAX_TOKEN_DURATION = 12
    TIME_RATE = 10.0 * 6 / 1000 / 3  #  3 times upsampled
    if len(us_alphas.shape) == 3:
        alphas, cif_peak = us_alphas[0], us_cif_peak[0]  # support inference batch_size=1 only
    if len(us_alphas.shape) == 2:
        alphas, peaks = us_alphas[0], us_peaks[0]  # support inference batch_size=1 only
    else:
        alphas, cif_peak = us_alphas, us_cif_peak
    num_frames = cif_peak.shape[0]
        alphas, peaks = us_alphas, us_peaks
    if char_list[-1] == '</s>':
        char_list = char_list[:-1]
    # char_list = [i for i in text]
    fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift  # total offset
    if len(fire_place) != len(char_list) + 1:
        alphas /= (alphas.sum() / (len(char_list) + 1))
        alphas = alphas.unsqueeze(0)
        peaks = cif_wo_hidden(alphas, threshold=1.0-1e-4)[0]
        fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift  # total offset
    num_frames = peaks.shape[0]
    timestamp_list = []
    new_char_list = []
    # for bicif model trained with large data, cif2 actually fires when a character starts
    # so treat the frames between two peaks as the duration of the former token
    fire_place = torch.where(cif_peak>1.0-1e-4)[0].cpu().numpy() - 1.5
    num_peak = len(fire_place)
    assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1
    fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift  # total offset
    # assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1
    # begin silence
    if fire_place[0] > START_END_THRESHOLD:
        char_list.insert(0, '<sil>')
        # char_list.insert(0, '<sil>')
        timestamp_list.append([0.0, fire_place[0]*TIME_RATE])
        new_char_list.append('<sil>')
    # tokens timestamp
    for i in range(len(fire_place)-1):
        # the peak is always a little ahead of the start time
        # timestamp_list.append([(fire_place[i]-1.2)*TIME_RATE, fire_place[i+1]*TIME_RATE])
        timestamp_list.append([(fire_place[i])*TIME_RATE, fire_place[i+1]*TIME_RATE])
        # cut the duration to token and sil of the 0-weight frames last long
        new_char_list.append(char_list[i])
        if MAX_TOKEN_DURATION < 0 or fire_place[i+1] - fire_place[i] <= MAX_TOKEN_DURATION:
            timestamp_list.append([fire_place[i]*TIME_RATE, fire_place[i+1]*TIME_RATE])
        else:
            # cut the duration to token and sil of the 0-weight frames last long
            _split = fire_place[i] + MAX_TOKEN_DURATION
            timestamp_list.append([fire_place[i]*TIME_RATE, _split*TIME_RATE])
            timestamp_list.append([_split*TIME_RATE, fire_place[i+1]*TIME_RATE])
            new_char_list.append('<sil>')
    # tail token and end silence
    # new_char_list.append(char_list[-1])
    if num_frames - fire_place[-1] > START_END_THRESHOLD:
        _end = (num_frames + fire_place[-1]) / 2
        _end = (num_frames + fire_place[-1]) * 0.5
        # _end = fire_place[-1] 
        timestamp_list[-1][1] = _end*TIME_RATE
        timestamp_list.append([_end*TIME_RATE, num_frames*TIME_RATE])
        char_list.append("<sil>")
        new_char_list.append("<sil>")
    else:
        timestamp_list[-1][1] = num_frames*TIME_RATE
    if begin_time:  # add offset time in model with vad
    if vad_offset:  # add offset time in model with vad
        for i in range(len(timestamp_list)):
            timestamp_list[i][0] = timestamp_list[i][0] + begin_time / 1000.0
            timestamp_list[i][1] = timestamp_list[i][1] + begin_time / 1000.0
            timestamp_list[i][0] = timestamp_list[i][0] + vad_offset / 1000.0
            timestamp_list[i][1] = timestamp_list[i][1] + vad_offset / 1000.0
    res_txt = ""
    for char, timestamp in zip(char_list, timestamp_list):
        res_txt += "{} {} {};".format(char, timestamp[0], timestamp[1])
    logging.warning(res_txt)  # for test
    for char, timestamp in zip(new_char_list, timestamp_list):
        #if char != '<sil>':
        if not sil_in_str and char == '<sil>': continue
        res_txt += "{} {} {};".format(char, str(timestamp[0]+0.0005)[:5], str(timestamp[1]+0.0005)[:5])
    res = []
    for char, timestamp in zip(char_list, timestamp_list):
    for char, timestamp in zip(new_char_list, timestamp_list):
        if char != '<sil>':
            res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)])
    return res_txt, res


def timestamp_sentence(punc_id_list, timestamp_postprocessed, text_postprocessed):
    punc_list = ['，', '。', '？', '、']
    res = []
    if text_postprocessed is None:
        return res
    if timestamp_postprocessed is None:
        return res
    if len(timestamp_postprocessed) == 0:
        return res
    if len(text_postprocessed) == 0:
        return res

    if punc_id_list is None or len(punc_id_list) == 0:
        res.append({
            'text': text_postprocessed.split(),
            "start": timestamp_postprocessed[0][0],
            "end": timestamp_postprocessed[-1][1],
            "timestamp": timestamp_postprocessed,
        })
        return res
    if len(punc_id_list) != len(timestamp_postprocessed):
        logging.warning("length mismatch between punc and timestamp")
    sentence_text = ""
    sentence_text_seg = ""
    ts_list = []
    sentence_start = timestamp_postprocessed[0][0]
    sentence_end = timestamp_postprocessed[0][1]
    texts = text_postprocessed.split()
    punc_stamp_text_list = list(zip_longest(punc_id_list, timestamp_postprocessed, texts, fillvalue=None))
    for punc_stamp_text in punc_stamp_text_list:
        punc_id, timestamp, text = punc_stamp_text
        # sentence_text += text if text is not None else ''
        if text is not None:
            if 'a' <= text[0] <= 'z' or 'A' <= text[0] <= 'Z':
                sentence_text += ' ' + text
            elif len(sentence_text) and ('a' <= sentence_text[-1] <= 'z' or 'A' <= sentence_text[-1] <= 'Z'):
                sentence_text += ' ' + text
            else:
                sentence_text += text
            sentence_text_seg += text + ' '
        ts_list.append(timestamp)

        punc_id = int(punc_id) if punc_id is not None else 1
        sentence_end = timestamp[1] if timestamp is not None else sentence_end

        if punc_id > 1:
            sentence_text += punc_list[punc_id - 2]
            res.append({
                'text': sentence_text,
                "start": sentence_start,
                "end": sentence_end,
                "timestamp": ts_list
            })
            sentence_text = ''
            sentence_text_seg = ''
            ts_list = []
            sentence_start = sentence_end
    return res