PUBLIC HARNESS BETA
We publish our numbers.
Most translation vendors ask you to trust a marketing demo. Pentecost publishes an open, reproducible benchmark measuring accuracy, latency, and cost across real audio datasets.
Methodology & Standard Metrics
Our benchmark harness evaluates live audio streams across speech datasets using independently verifiable metrics: Word Error Rate (WER) for transcription precision, BLEU and COMET scores for translation quality, p95 caption latency (95th percentile delay), and cost per minute.
Pentecost Pipeline Quality Summary (Q2 2026 Audit)
| Metric | Pentecost Speech Pipeline | Target Bar |
|---|---|---|
| Caption Latency (p95) | 1.28 seconds | < 1.50 seconds |
| Transcription WER (English Speech Precision) | 3.82% error (96.18% accuracy) | < 5.00% error |
| Domain Keyword Hit Rate | 98.4% hit rate | > 95.0% |
| Hallucination / Contamination Guard | 0.02% phantom rate | < 0.10% |