{
  "@context": "https://schema.org",
  "@type": "Dataset",
  "name": "LongScribe cross-vendor transcription holdout-v2",
  "description": "Frozen 20-sample transcription holdout comparing complete LongScribe High Accuracy, Notta and Descript runs, with incomplete provider runs labeled separately.",
  "url": "https://longscribe.com/best-ai-transcription-software",
  "contentUrl": "https://longscribe.com/data/longscribe-transcription-holdout-v2.json",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "datePublished": "2026-09-13",
  "dateModified": "2026-09-19",
  "creator": {
    "@type": "Organization",
    "name": "LongScribe",
    "url": "https://longscribe.com/"
  },
  "benchmark_id": "longscribe-transcription-holdout-v2",
  "dataset_manifest_sha256": "12ca6ba0a8ff38c5c4fc96b2422b44343149d5d54a5aadfec5696d2b0dcc2703",
  "normalization_version": "asr-text-v1",
  "sample_count": 20,
  "reference_words": 540,
  "full_runs": [
    {
      "product": "LongScribe High Accuracy",
      "engine": "ElevenLabs Scribe v2",
      "tested_at": "2026-09-13",
      "completed_samples": 20,
      "total_samples": 20,
      "word_errors": 77,
      "reference_words": 540,
      "word_error_rate": 0.1425925926,
      "character_error_rate": 0.085046729,
      "ranking_eligible": true
    },
    {
      "product": "Notta",
      "tested_at": "2026-09-13",
      "completed_samples": 20,
      "total_samples": 20,
      "word_errors": 96,
      "reference_words": 540,
      "word_error_rate": 0.1777777778,
      "character_error_rate": 0.1172897,
      "ranking_eligible": true
    },
    {
      "product": "Descript",
      "mode": "default automatic transcription",
      "tested_at": "2026-09-13",
      "completed_samples": 20,
      "total_samples": 20,
      "word_errors": 108,
      "reference_words": 540,
      "word_error_rate": 0.2,
      "character_error_rate": 0.124766,
      "ranking_eligible": true
    }
  ],
  "partial_runs": [
    {
      "product": "TurboScribe",
      "tested_at": "2026-09-19",
      "completed_samples": 12,
      "total_samples": 20,
      "word_errors": 23,
      "reference_words": 278,
      "word_error_rate": 0.0827338129,
      "character_error_rate": 0.0462,
      "critical_term_recall": 0.9444,
      "ranking_eligible": false,
      "status": "preliminary_incomplete"
    }
  ],
  "separate_benchmarks": [
    {
      "product": "Happy Scribe",
      "benchmark_id": "longscribe-transcription-benchmark-v1",
      "completed_samples": 20,
      "total_samples": 20,
      "word_error_rate": 0.2033,
      "character_error_rate": 0.1163,
      "critical_term_recall": 0.9,
      "note": "Complete result on benchmark-v1; not directly merged with holdout-v2 scores."
    }
  ],
  "limitations": [
    "This holdout is short-clip product-selection evidence, not proof of universal accuracy across all recordings or languages.",
    "The holdout became development evidence after its results were used to select the High Accuracy path; future model or routing changes require a fresh independent holdout.",
    "TurboScribe is incomplete at 12/20 samples and is excluded from the complete 20/20 comparison.",
    "Happy Scribe's complete 20/20 result belongs to benchmark-v1, a different frozen dataset, and is reported separately."
  ]
}