paperdatasfiltered1/dat_8_281h

Private Silero-filtered copy of the source dataset.

• Source dataset: linear-schedules-experiment2/dat_8_285h
• Source revision: 7284f1f53135b6f55ca912f8eccb7ed9cd1912d2
• Input rows: 64393
• Retained rows: 63482
• Removed rows: 911
• The original columns, audio payloads, metadata, and source row order are preserved.
• Only rows whose Silero mapping has keep=true are included.

Silero VAD used the 16 kHz ONNX model at commit fba061d (silero_vad_16k_op15.onnx, sha256 794ed8a51d4f37faf0555383aa34dbaeeb83e3031a1df1e0351c457e1142bd3e); the speech-start threshold was 0.5, speech-stop threshold 0.35, minimum silence 0.10 s, speech padding 0.03 s, and minimum qualifying region 0.25 s. keep=true means at least one qualifying region was found. The audio was decoded with ffmpeg to mono 16 kHz.

Filtering provenance: paperdatasfiltered1/dat_8_281h/vad_filter_provenance.json; per-row decisions: vad_mapping.parquet.