{
  "baseline": {
    "examples": 128,
    "reference_words": 1466,
    "word_errors": 3161,
    "wer": 2.1562073669849933,
    "cer": 1.5107363983664865,
    "raw_wer": 2.167912984364378,
    "median_inference_seconds": 0.14351516700116917,
    "p95_inference_seconds": 1.2805451527354308,
    "total_inference_seconds": 46.68536271119956,
    "audio_seconds": 770.861,
    "aggregate_real_time_factor": 0.06056262116153179,
    "median_output_tokens": 16.0,
    "generation_cap_count": 26
  },
  "fine_tuned": {
    "examples": 128,
    "reference_words": 1466,
    "word_errors": 1133,
    "wer": 0.7728512960436562,
    "cer": 0.5401132920563826,
    "raw_wer": 0.7763426240652618,
    "median_inference_seconds": 0.33312687498982996,
    "p95_inference_seconds": 0.6850233413453678,
    "total_inference_seconds": 46.22849254927132,
    "audio_seconds": 770.861,
    "aggregate_real_time_factor": 0.05996994600747907,
    "median_output_tokens": 45.5,
    "generation_cap_count": 1
  },
  "absolute_wer_reduction_percentage_points": 138.3356070941337,
  "relative_wer_reduction_percent": 64.15691236950332,
  "median_inference_time_change_percent": 132.11963024585134,
  "timing_method": "batch 1; fp32; same GPU; two validation warmups/model; alternating model order; explicit device sync; cached log-Mel inputs; host-to-device transfer plus greedy generation; excludes audio loading/feature extraction/model load/network",
  "limitations": [
    "Public lecture subset, not proprietary Zoice calls or telephony.",
    "Single seed and small held-out subset; results are illustrative.",
    "Timing depends on output length; no claim of kernel or streaming optimization.",
    "Dataset speaker_id/session disjointness does not verify real-world speaker identity.",
    "Public data may overlap unknown model pretraining; pretraining overlap was not audited."
  ],
  "token_metadata_note": "Token counts and cap flags corrected in an identical-hypothesis replay; original per-clip timings retained. See token_metadata_audit.json."
}