window size

[Jundo26]

why max window size == 500ms ？
Is it because the duration of a word is about 500ms?

[ljj7975]

I think max_window_size can be a misleading name.
It is prefixed with max as samples can possibly have variable length (shorter than the max_window_size)

The window is the single unit of a sample that is preprocessed together and fed into the model
The following code should be self-explanatory

howl/howl/model/inference.py

Lines 169 to 193 in 55f8b92

	def infer(self, audio_data: torch.Tensor) -> bool:
	sequence_present = False
	for window in stride(audio_data, self.max_window_size_ms, self.eval_stride_size_ms, self.sample_rate):
	if window.size(-1) < 1000:
	break
	self.ingest_frame(window.squeeze(0), curr_time=self.curr_time)
	self.curr_time += self.eval_stride_size_ms
	if self.sequence_present(self.curr_time):
	sequence_present = True
	break
	return sequence_present
	@torch.no_grad()
	def ingest_frame(self, x: torch.Tensor, lengths: torch.Tensor = None, curr_time: float = None) -> int:
	self.std = self.std.to(x.device)
	if lengths is None:
	lengths = torch.tensor([x.size(-1)]).to(x.device)
	lengths = self.std.compute_lengths(lengths)
	x = self.zmuv(self.std(x.unsqueeze(0)))
	p = self.model(x, lengths).softmax(-1)[0].cpu().numpy()
	p *= self.inference_weights
	p = p / p.sum()
	label = self._append_probability_frame(p, curr_time=curr_time)
	return label