Detecting the voices with low energy

[v-yunbin]

For relatively small sounds, it cannot be detected. Can these small sounds be detected by adjusting the parameters? For these small sounds, I can do it via webrtc vad mode=2.

[snakers4]

If small means quiet, then you may plot the probability plot (visualize_probs=True), and tweak the most obvious params here:

silero-vad/utils_vad.py

Lines 164 to 175 in 563106e

	def get_speech_timestamps(audio: torch.Tensor,
	model,
	threshold: float = 0.5,
	sampling_rate: int = 16000,
	min_speech_duration_ms: int = 250,
	max_speech_duration_s: float = float('inf'),
	min_silence_duration_ms: int = 100,
	window_size_samples: int = 512,
	speech_pad_ms: int = 30,
	return_seconds: bool = False,
	visualize_probs: bool = False,
	progress_tracking_callback: Callable[[float], None] = None):

[v-yunbin]

the small not means quiet and the voices are small but sounds clear. it doesn't seem to work for weaking the most obvious params.