v0.8.0

游雁

2023-10-10 580b11b57ac4b62f7e2acda73813a4e10e8e4cd3

 funasr/models/e2e_vad.py

@@ -5,6 +5,7 @@
from torch import nn
import math
from funasr.models.encoder.fsmn_encoder import FSMN
from funasr.models.base_model import FunASRModel


class VadStateMachine(Enum):
@@ -211,7 +212,7 @@
        return int(self.frame_size_ms)


class E2EVadModel(nn.Module):
class E2EVadModel(FunASRModel):
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Deep-FSMN for Large Vocabulary Continuous Speech Recognition
@@ -296,13 +297,14 @@
        self.sil_frame = 0
        self.frame_probs = []

        assert self.output_data_buf[-1].contain_seg_end_point == True
        drop_frames = int(self.output_data_buf[-1].end_ms / self.vad_opts.frame_in_ms)
        real_drop_frames = drop_frames - self.last_drop_frames
        self.last_drop_frames = drop_frames
        self.data_buf_all = self.data_buf_all[real_drop_frames * int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):]
        self.decibel = self.decibel[real_drop_frames:]
        self.scores = self.scores[:, real_drop_frames:, :]
        if self.output_data_buf:
            assert self.output_data_buf[-1].contain_seg_end_point == True
            drop_frames = int(self.output_data_buf[-1].end_ms / self.vad_opts.frame_in_ms)
            real_drop_frames = drop_frames - self.last_drop_frames
            self.last_drop_frames = drop_frames
            self.data_buf_all = self.data_buf_all[real_drop_frames * int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):]
            self.decibel = self.decibel[real_drop_frames:]
            self.scores = self.scores[:, real_drop_frames:, :]

    def ComputeDecibel(self) -> None:
        frame_sample_length = int(self.vad_opts.frame_length_ms * self.vad_opts.sample_rate / 1000)