funasr1.0

游雁

2024-01-13 bdfd27b9e96bd55c449953bb577e1d4deeaf11c9

 funasr/models/monotonic_aligner/model.py

@@ -4,15 +4,16 @@
from torch.cuda.amp import autocast
from typing import Union, Dict, List, Tuple, Optional

from funasr.register import tables
from funasr.utils import postprocess_utils
from funasr.utils.datadir_writer import DatadirWriter
from funasr.models.paraformer.cif_predictor import mae_loss
from funasr.train_utils.device_funcs import force_gatherable
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
from funasr.models.transformer.utils.nets_utils import make_pad_mask
from funasr.utils.timestamp_tools import ts_prediction_lfr6_standard
from funasr.utils.load_utils import load_audio_and_text_image_video, extract_fbank, load_audio_and_text_image_video
from funasr.utils import postprocess_utils
from funasr.utils.datadir_writer import DatadirWriter
from funasr.register import tables
from funasr.models.ctc.ctc import CTC
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank


@tables.register("model_classes", "monotonicaligner")
@@ -153,7 +154,7 @@
        meta_data = {}
        # extract fbank feats
        time1 = time.perf_counter()
        audio_list, text_token_int_list = load_audio_and_text_image_video(data_in, 
        audio_list, text_token_int_list = load_audio_text_image_video(data_in, 
                                                                            fs=frontend.fs, 
                                                                            audio_fs=kwargs.get("fs", 16000), 
                                                                            data_type=kwargs.get("data_type", "sound"), 
@@ -165,7 +166,8 @@
        meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
        meta_data["batch_data_time"] = speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            
        speech.to(device=kwargs["device"]), speech_lengths.to(device=kwargs["device"])
        speech = speech.to(device=kwargs["device"])
        speech_lengths = speech_lengths.to(device=kwargs["device"])

        # Encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
@@ -187,9 +189,12 @@
            text_postprocessed, time_stamp_postprocessed, _ = postprocess_utils.sentence_postprocess(token, timestamp)
            result_i = {"key": key[i], "text": text_postprocessed,
                                "timestamp": time_stamp_postprocessed,
                                }    
            # ibest_writer["token"][key[i]] = " ".join(token)
            ibest_writer["timestamp_list"][key[i]] = time_stamp_postprocessed
            ibest_writer["timestamp_str"][key[i]] = timestamp_str
                                }
            results.append(result_i)

            if ibest_writer:
                # ibest_writer["token"][key[i]] = " ".join(token)
                ibest_writer["timestamp_list"][key[i]] = time_stamp_postprocessed
                ibest_writer["timestamp_str"][key[i]] = timestamp_str
            
        return results, meta_data