Merge pull request #473 from alibaba-damo-academy/dev_smohan

jmwang66

2023-05-09 8dab6d184a034ca86eafa644ea0d2100aadfe27d

 funasr/models/e2e_vad.py

@@ -35,6 +35,11 @@


class VADXOptions:
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Deep-FSMN for Large Vocabulary Continuous Speech Recognition
    https://arxiv.org/abs/1803.05030
    """
    def __init__(
            self,
            sample_rate: int = 16000,
@@ -99,6 +104,11 @@


class E2EVadSpeechBufWithDoa(object):
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Deep-FSMN for Large Vocabulary Continuous Speech Recognition
    https://arxiv.org/abs/1803.05030
    """
    def __init__(self):
        self.start_ms = 0
        self.end_ms = 0
@@ -117,6 +127,11 @@


class E2EVadFrameProb(object):
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Deep-FSMN for Large Vocabulary Continuous Speech Recognition
    https://arxiv.org/abs/1803.05030
    """
    def __init__(self):
        self.noise_prob = 0.0
        self.speech_prob = 0.0
@@ -126,6 +141,11 @@


class WindowDetector(object):
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Deep-FSMN for Large Vocabulary Continuous Speech Recognition
    https://arxiv.org/abs/1803.05030
    """
    def __init__(self, window_size_ms: int, sil_to_speech_time: int,
                 speech_to_sil_time: int, frame_size_ms: int):
        self.window_size_ms = window_size_ms
@@ -192,6 +212,11 @@


class E2EVadModel(nn.Module):
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Deep-FSMN for Large Vocabulary Continuous Speech Recognition
    https://arxiv.org/abs/1803.05030
    """
    def __init__(self, encoder: FSMN, vad_post_args: Dict[str, Any], frontend=None):
        super(E2EVadModel, self).__init__()
        self.vad_opts = VADXOptions(**vad_post_args)
@@ -286,7 +311,7 @@
                                0.000001))

    def ComputeScores(self, feats: torch.Tensor, in_cache: Dict[str, torch.Tensor]) -> None:
        scores = self.encoder(feats, in_cache)  # return B * T * D
        scores = self.encoder(feats, in_cache).to('cpu')  # return B * T * D
        assert scores.shape[1] == feats.shape[1], "The shape between feats and scores does not match"
        self.vad_opts.nn_eval_block_size = scores.shape[1]
        self.frm_cnt += scores.shape[1]  # count total frames