Merge pull request #177 from alibaba-damo-academy/dev_timestamp

zhifu gao

2023-03-02 94cb66dbb9ae12e044a41fb8a3d84e1835ee7e7b

Merge pull request #177 from alibaba-damo-academy/dev_timestamp

Dev timestamp

 funasr/runtime/python/onnxruntime/demo.py

@@ -2,7 +2,8 @@
from rapid_paraformer import Paraformer

model_dir = "/Users/shixian/code/funasr2/export/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
model_dir = "/Users/shixian/code/funasr2/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
# model_dir = "/Users/shixian/code/funasr2/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"

model = Paraformer(model_dir, batch_size=1)

wav_path = ['/Users/shixian/code/funasr2/export/damo/speech_paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch/example/asr_example.wav']

 funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py

@@ -14,7 +14,7 @@
                          read_yaml)
from .utils.postprocess_utils import sentence_postprocess
from .utils.frontend import WavFrontend
from funasr.utils.timestamp_tools import time_stamp_lfr6_pl
from .utils.timestamp_utils import time_stamp_lfr6_onnx

logging = get_logger()

@@ -41,17 +41,16 @@
        )
        self.ort_infer = OrtInferSession(model_file, device_id)
        self.batch_size = batch_size
        self.plot = True

    def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:
        waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
        waveform_nums = len(waveform_list)

        asr_res = []
        for beg_idx in range(0, waveform_nums, self.batch_size):
            res = {}
            end_idx = min(waveform_nums, beg_idx + self.batch_size)
            feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])

            try:
                outputs = self.infer(feats, feats_len)
                am_scores, valid_token_lens = outputs[0], outputs[1]
@@ -68,11 +67,17 @@
                preds, raw_token = self.decode(am_scores, valid_token_lens)[0]
                res['preds'] = preds
                if us_cif_peak is not None:
                    timestamp = time_stamp_lfr6_pl(us_alphas, us_cif_peak, copy.copy(raw_token), log=False)
                    timestamp, timestamp_total = time_stamp_lfr6_onnx(us_cif_peak, copy.copy(raw_token))
                    res['timestamp'] = timestamp
                    if self.plot:
                        self.plot_wave_timestamp(waveform_list[0], timestamp_total)
            asr_res.append(res)
        return asr_res

    def plot_wave_timestamp(self, wav, text_timestamp):
        # TODO: Plot the wav and timestamp results with matplotlib
        import pdb; pdb.set_trace()

    def load_data(self,
                  wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List:
        def load_wav(path: str) -> np.ndarray:

 funasr/runtime/python/onnxruntime/rapid_paraformer/utils/timestamp_utils.py

New file
@@ -0,0 +1,58 @@
import numpy as np


def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0):
    if not len(char_list):
        return []
    START_END_THRESHOLD = 5
    MAX_TOKEN_DURATION = 14
    TIME_RATE = 10.0 * 6 / 1000 / 3  #  3 times upsampled
    cif_peak = us_cif_peak.reshape(-1)
    num_frames = cif_peak.shape[-1]
    if char_list[-1] == '</s>':
        char_list = char_list[:-1]
    # char_list = [i for i in text]
    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 = np.where(cif_peak>1.0-1e-4)[0] - 1.5  # np format
    num_peak = len(fire_place)
    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>')
        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):
        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
    if num_frames - fire_place[-1] > START_END_THRESHOLD:
        _end = (num_frames + fire_place[-1]) / 2
        timestamp_list[-1][1] = _end*TIME_RATE
        timestamp_list.append([_end*TIME_RATE, num_frames*TIME_RATE])
        new_char_list.append("<sil>")
    else:
        timestamp_list[-1][1] = num_frames*TIME_RATE
    if begin_time:  # 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
    assert len(new_char_list) == len(timestamp_list)
    res_total = []
    for char, timestamp in zip(new_char_list, timestamp_list):
        res_total.append([char, timestamp[0], timestamp[1]])  # += "{} {} {};".format(char, timestamp[0], timestamp[1])
    res = []
    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, res_total

 funasr/utils/timestamp_tools.py

@@ -55,6 +55,7 @@
            res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)])
    return res


def time_stamp_sentence(punc_id_list, time_stamp_postprocessed, text_postprocessed):
    res = []
    if text_postprocessed is None: