Bug report — Incorrectly Detects Blowing into Mic as Voice

[s1lviu]

The Silero Voice Activity Detection (VAD) model incorrectly classifies the sound of blowing directly into the microphone as human speech. This issue persists across various blowing intensities, from soft to strong. The expected behavior is for the VAD to recognize these non-speech, noise-like sounds as non-voice activity.

Reproduction Steps

Web Application Test:

• Navigate to the web-based VAD tool at https://www.vad.ricky0123.com/.
• Use the tool to record the sound of blowing into the microphone.
• Observe the VAD output.

C++ Example Test:

• Clone the Silero VAD repository from https://github.com/snakers4/silero-vad/tree/master/examples/cpp.
• Compile and run the C++ example according to the provided instructions.
• Input audio by blowing into the microphone.
• Observe the VAD output.

Expected Behavior
The VAD should classify the blowing sounds as non-voice activity, distinguishing them from human speech.

Actual Behavior
The VAD system classifies the blowing sounds as voice activity, indicating a false positive in voice detection.

[DanyPell]

Do you have the probability number?

[s1lviu]

Do you have the probability number?

Around 65-70%.

[lionsheep0724]

same here. ringback tone detected as voice with 94% of probability.