{
  "dataset": "VoxLingua107",
  "split": "dev",
  "utterances": 1609,
  "languages": 33,
  "metric": "utterance-level accuracy (%)",
  "entrypoint": "espnet2.bin.lid_inference",
  "implementation_revision": "ca777d59eb49519e3827e40cc34997f2123d6682",
  "device": "cuda",
  "dtype": "float32",
  "batch_size": 1,
  "num_workers": 8,
  "use_preprocessor": false,
  "fix_duration": false,
  "input_order": "wav.scp sorted by utterance ID; standard ESPnet streaming iterator",
  "manifest_sha256": "6ad3c8fc6c74fab2720fe186525fc3bd20c4ab7b1be35d227c5c7f5456ba6c54",
  "environment": {
    "python": "3.12.3 (main, Jun 19 2026, 12:46:00) [GCC 13.3.0]",
    "torch": "2.11.0+cu128",
    "torch_cuda_build": "12.8",
    "torchaudio": "2.11.0+cu128",
    "transformers": "5.14.1",
    "espnet_s3prl": "2026.5.14",
    "soundfile": "0.14.0"
  },
  "repo_id": "espnet/geolid_vl107only_independent_trainable",
  "model_sha256": "a4a77caebec41b539899c2760b1216c6c6d68e3d5093b1095ed07b1e7f8a0064",
  "correct": 1509,
  "accuracy_percent": 93.78495960223742,
  "paper_accuracy_percent": 93.7,
  "source_hf_revision_before_layout_migration": "71c1ec0f546c65eddfaf45584042b9e15b0278aa",
  "batch_dependence": {
    "s3prl_frame_alignment": "S3PRL reports ceil(samples/320) frames and repeats the final frame only at the batch boundary. Shorter items can instead expose padded-query frames from Wav2Vec2.",
    "ecapa_tdnn": "Temporal convolutions and unmasked SE global means depend on the padded sequence length before the masked final pooling.",
    "mvn_and_final_pooling_mask_padding": true,
    "reported_measurement_batch_size": 1
  }
}