Skip to main content

Fixed an issue where including "16k" in the path caused the default ONNX model to be mistakenly identified as 16 kHz-only

silero-vad: Issue where a 16k in the path causes a standard ONNX model to be mistaken for a 16 kHz-specific model​

Link: https://github.com/snakers4/silero-vad/pull/814

Background​

OnnxWrapper checks for '16k' in path to distinguish 16 kHz-specific models (silero_vad_16k_op15.onnx). However, path refers to the full path, not just the filename. Since load_silero_vad(onnx=True) passes the internal path of the installed package, if 16k is included in the name of a virtual environment or project directory (e.g., asr_16k_project/.venv/...), even the default model silero_vad.onnx—which supports both 8 kHz and 16 kHz—is treated as 16 kHz-only.

Cause​

In src/silero_vad/utils_vad.py’s OnnxWrapper.__init__ (line 79 in the master branch), a substring check is performed on the full path. If a match is found, sample_rates is set to [16000] and a warning is output; subsequently, 8 kHz input is rejected in _validate_input and set to ValueError: Supported sampling rates: [16000] (or multiply of 16000).

Steps to Reproduce​

Copy the default model file to a directory named asr_16k_project, load it as OnnxWrapper, and then call get_speech_timestamps with 8 kHz audio.

original location : [8000, 16000] []
asr_16k_project/  : [16000] ['This model support only 16000 sampling rate!']
ValueError: Supported sampling rates: [16000] (or multiply of 16000)

Changes​

  • Limited the scope of the check to os.path.basename(path) (added 2 lines, including the addition of import os; deleted 1 line)
  • New test: Verify that 8 kHz works even if 16k exists in the parent directory, and that the op15 model is still recognized as 16 kHz-only

Verification​

  • Before the fix: 1 new test failed (due to the warning above); after the fix: 2 tests passed
  • pytest tests overall: 47 passed, 10 skipped (skipped due to torchcodec not being installed)
  • Windows paths cannot be executed locally