| | |
|---|
| | | from pathlib import Path |
|---|
| | | from typing import List, Union, Tuple |
|---|
| | | |
|---|
| | | import copy |
|---|
| | | import librosa |
|---|
| | | import numpy as np |
|---|
| | | |
|---|
| | |
|---|
| | | read_yaml) |
|---|
| | | from .utils.postprocess_utils import sentence_postprocess |
|---|
| | | from .utils.frontend import WavFrontend |
|---|
| | | from .utils.timestamp_utils import time_stamp_lfr6_onnx |
|---|
| | | |
|---|
| | | logging = get_logger() |
|---|
| | | |
|---|
| | |
|---|
| | | ) |
|---|
| | | 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: |
|---|
| | | am_scores, valid_token_lens = self.infer(feats, feats_len) |
|---|
| | | outputs = self.infer(feats, feats_len) |
|---|
| | | am_scores, valid_token_lens = outputs[0], outputs[1] |
|---|
| | | if len(outputs) == 4: |
|---|
| | | # for BiCifParaformer Inference |
|---|
| | | us_alphas, us_cif_peak = outputs[2], outputs[3] |
|---|
| | | else: |
|---|
| | | us_alphas, us_cif_peak = None, None |
|---|
| | | except ONNXRuntimeError: |
|---|
| | | #logging.warning(traceback.format_exc()) |
|---|
| | | logging.warning("input wav is silence or noise") |
|---|
| | | preds = [''] |
|---|
| | | else: |
|---|
| | | preds = self.decode(am_scores, valid_token_lens) |
|---|
| | | |
|---|
| | | asr_res.extend(preds) |
|---|
| | | preds, raw_token = self.decode(am_scores, valid_token_lens)[0] |
|---|
| | | res['preds'] = preds |
|---|
| | | if us_cif_peak is not None: |
|---|
| | | 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 infer(self, feats: np.ndarray, |
|---|
| | | feats_len: np.ndarray) -> Tuple[np.ndarray, np.ndarray]: |
|---|
| | | am_scores, token_nums = self.ort_infer([feats, feats_len]) |
|---|
| | | return am_scores, token_nums |
|---|
| | | outputs = self.ort_infer([feats, feats_len]) |
|---|
| | | return outputs |
|---|
| | | |
|---|
| | | def decode(self, am_scores: np.ndarray, token_nums: int) -> List[str]: |
|---|
| | | return [self.decode_one(am_score, token_num) |
|---|
| | |
|---|
| | | |
|---|
| | | # Change integer-ids to tokens |
|---|
| | | token = self.converter.ids2tokens(token_int) |
|---|
| | | token = token[:valid_token_num-1] |
|---|
| | | # token = token[:valid_token_num-1] |
|---|
| | | texts = sentence_postprocess(token) |
|---|
| | | text = texts[0] |
|---|
| | | # text = self.tokenizer.tokens2text(token) |
|---|
| | | return text |
|---|
| | | return text, token |
|---|
| | | |
|---|
