{
  "benchmark": "Simple Speech Eval",
  "title": "Rime Prism and Phonic join our voice benchmark",
  "cases": 10,
  "repetitions": 3,
  "interval": "case-clustered bootstrap 95%; 2000 resamples; seed 20260925; unknown criteria at midpoint",
  "caveat": "Results from different dates are separate cohorts and were not rerun. Caller voices, time limits and assessment versions differ between dates. Ten scenarios are a small sample.",
  "chart_notes": [
    "Dot: p50 response time · horizontal whisker: p90 · vertical bar: 95% bootstrap interval on the score",
    "Results from different dates were not rerun. Caller voices, time limits and assessment versions differ.",
    "Response time is word-aligned on the simulated caller’s recording: last caller word to first reply word.",
    "Dialt Smart’s zero-width interval reflects this sample only."
  ],
  "latency": "Word-aligned response time: last caller word end to first reply word start, CPU forced alignment. p50 and p90 over aligned responses pooled across the 30 conversations. Every model is measured on the simulated caller recording.",
  "arms": [
    {"label": "Dialt Fast", "group": "dialt", "date": "2026-11-11", "date_label": "11 Nov 2026", "score": 0.972223, "interval": [0.9417, 1.0], "attempts": 30, "passed": 27, "failed": 3, "inconclusive": 0, "errors": 0, "caller": "US Male", "limit_s": 120, "p50_ms": 1291.5, "p90_ms": 2064, "aligned_responses": 90, "clock": "caller recording"},
    {"label": "Rime Prism", "group": "new", "date": "2026-11-12", "date_label": "12 Nov 2026", "score": 0.845277, "interval": [0.7011, 0.9667], "attempts": 30, "passed": 17, "failed": 13, "inconclusive": 0, "errors": 0, "caller": "US Male", "limit_s": 120, "note": "Submitted one conversation at a time because the account admits one concurrent session.", "p50_ms": 1197, "p90_ms": 1747, "aligned_responses": 91, "clock": "caller recording"},
    {"label": "Phonic", "group": "new", "date": "2026-11-12", "date_label": "12 Nov 2026", "score": 0.886117, "interval": [0.8, 0.9667], "attempts": 30, "passed": 20, "failed": 9, "inconclusive": 1, "errors": 0, "caller": "US Male", "limit_s": 120, "note": "Word-by-word recognition is grouped into one utterance per caller turn. The inconclusive conversation is one where the simulated caller did not deliver its scripted correction; its assessed criteria count toward the score.", "p50_ms": 1431, "p90_ms": 1628, "aligned_responses": 96, "clock": "caller recording"},
    {"label": "Dialt Smart", "group": "dialt", "date": "2026-10-21", "date_label": "21 Oct 2026", "score": 1, "interval": [1, 1], "attempts": 30, "p50_ms": 1922, "p90_ms": 3335, "aligned_responses": 94, "clock": "caller recording"},
    {"label": "Gemini 3.8 Live", "group": "earlier", "date": "2026-09-26", "date_label": "26 Sep 2026", "score": 0.899, "interval": [0.8097, 0.9778], "attempts": 30, "caller": "Irish Male", "limit_s": 120, "p50_ms": 2018, "p90_ms": 2720, "aligned_responses": 84, "clock": "caller recording"},
    {"label": "Eleven Realtime Max", "group": "earlier", "date": "2026-10-21", "date_label": "21 Oct 2026", "score": 0.876852, "interval": [0.8125, 0.9338], "attempts": 30, "caller": "US Male", "limit_s": 180, "p50_ms": 2094, "p90_ms": 3395, "aligned_responses": 90, "clock": "caller recording"},
    {"label": "GPT Live-1", "group": "earlier", "date": "2026-10-21", "date_label": "21 Oct 2026", "score": 0.7967595, "interval": [0.6653, 0.9148], "attempts": 30, "caller": "US Male", "limit_s": 180, "p50_ms": 2273, "p90_ms": 3225, "aligned_responses": 94, "clock": "caller recording"},
    {"label": "Eleven Realtime Mini", "group": "earlier", "date": "2026-10-21", "date_label": "21 Oct 2026", "score": 0.709074, "interval": [0.6014, 0.8208], "attempts": 30, "caller": "US Male", "limit_s": 180, "p50_ms": 2104.5, "p90_ms": 3544, "aligned_responses": 100, "clock": "caller recording"}
  ],
  "superseded": [
    {"label": "Dialt Fast", "date": "2026-10-21", "score": 0.942222, "interval": [0.8956, 0.9867], "attempts": 30, "note": "Replaced on the chart by the 11 November run.", "p50_ms": 1513, "p90_ms": 2748, "aligned_responses": 95, "clock": "caller recording"}
  ],
  "not_assessed": [
    {"label": "Sophon", "reason": "Its integration returns no transcript of the caller speech it received, which every scenario in this suite needs to confirm what the model heard. No conversations were run."}
  ],
  "earlier_sources": "../on-prem/data.json"
}
