#Loop over batches and time steps
        for batch in range(vad_th.shape[0]):
            for time_step in range(vad_th.shape[1] - 1):
                if (
                    vad_th[batch, time_step] == 2
                    and vad_th[batch, time_step + 1] == 1
                ):
                    vad_th[batch, time_step + 1] = 2

vad_th = vad_th.cpu().numpy()
[...]
vad_th = torch.from_numpy(vad_th)

    def apply_threshold(
        self, vad_prob, activation_th=0.5, deactivation_th=0.25
    ):
        """Scans the frame-level speech probabilities and applies a threshold
        on them. Speech starts when a value larger than activation_th is
        detected, while it ends when observing a value lower than
        the deactivation_th.

        Arguments
        ---------
        vad_prob: torch.Tensor
            Frame-level speech probabilities.
        activation_th:  float
            Threshold for starting a speech segment.
        deactivation_th: float
            Threshold for ending a speech segment.

        Returns
        -------
        vad_th: torch.Tensor
            Tensor containing 1 for speech regions and 0 for non-speech regions.
        """
        vad_activation = (vad_prob >= activation_th).int()
        vad_deactivation = (vad_prob >= deactivation_th).int()
        vad_th = vad_activation + vad_deactivation

        # Move tensor to cpu, make numpy array
        vad_th = vad_th.cpu().numpy()

        #Loop over batches and time steps
        for batch in range(vad_th.shape[0]):
            for time_step in range(vad_th.shape[1] - 1):
                if (
                    vad_th[batch, time_step] == 2
                    and vad_th[batch, time_step + 1] == 1
                ):
                    vad_th[batch, time_step + 1] = 2

        # Get tensor from numpy
        vad_th = torch.from_numpy(vad_th)
        vad_th[vad_th == 1] = 0
        vad_th[vad_th == 2] = 1
        return vad_th

from speechbrain.inference.VAD import VAD

from time import perf_counter

VAD = VAD.from_hparams(
    source="speechbrain/vad-crdnn-libriparty",
    savedir="pretrained_models/vad-crdnn-libriparty",
    huggingface_cache_dir="./",
    run_opts={"device": "cuda"}
)

s = perf_counter()
boundaries = VAD.get_speech_segments(
    "path_to_my_6h_audio.wav"
)
e = perf_counter()

# Print the output
VAD.save_boundaries(boundaries)
print(f"It took {e-s}")