update

speech_asr

2023-03-15 7c6ed3830acf6413ab86fd9a5f38825db617f989

update

 funasr/bin/eend_ola_inference.py

@@ -17,6 +17,7 @@
import numpy as np
import torch
from typeguard import check_argument_types
from scipy.signal import medfilt

from funasr.models.frontend.wav_frontend import WavFrontendMel23
from funasr.tasks.diar import EENDOLADiarTask
@@ -234,9 +235,22 @@
            # batch = {k: v[0] for k, v in batch.items() if not k.endswith("_lengths")}

            results = speech2diar(**batch)

            # post process
            a = medfilt(results[0], (11, 1))
            rst = []
            for spkid, frames in enumerate(a.T):
                frames = np.pad(frames, (1, 1), 'constant')
                changes, = np.where(np.diff(frames, axis=0) != 0)
                fmt = "SPEAKER {:s} 1 {:7.2f} {:7.2f} <NA> <NA> {:s} <NA>"
                for s, e in zip(changes[::2], changes[1::2]):
                    st = s / 10.
                    ed = e / 10.
                    rst.append(fmt.format(keys[0], st, ed, "{}_{}".format(keys[0],str(spkid))))

            # Only supporting batch_size==1
            key, value = keys[0], output_results_str(results, keys[0])
            item = {"key": key, "value": value}
            value = "\n".join(rst)
            item = {"key": keys[0], "value": value}
            result_list.append(item)
            if output_path is not None:
                output_writer.write(value)