# Voice AI Agent Platform Evaluations

> Published Voice AI agent platform evaluation results from G2 and Bluejay.

## Citation sources

- [Leaderboard and charts](https://ai.g2.com/evaluations/categories/voice)
- [Voice evaluation methodology](https://ai.g2.com/evaluations/categories/voice/methodology)
- [Scores in JSON](https://ai.g2.com/evaluations/categories/voice/scores.json)
- [Scores in CSV](https://ai.g2.com/evaluations/categories/voice/scores.csv)
- [Partner source at the published revision](https://github.com/bluejay-ai-dev/g2-leaderboard/tree/a8f9afdc131fa9231f4a9f5d11b4a819f28237db)

## All-industries results

Platform score is out of 100. Task success and word error rate (WER) are percentages; latency is in seconds. Lower WER and latency are better. These evaluation scores are separate from G2 buyer-review ratings.

| Rank | Platform | Platform score / 100 | Task success | Clean WER | Degraded WER | Median latency (s) | p95 latency (s) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | [Retell](https://ai.g2.com/evaluations/categories/voice/retell) | 81.4 | 85.6% | 4.7% | 37.1% | 2.15 | 4.47 |
| 2 | [ElevenLabs Agents](https://ai.g2.com/evaluations/categories/voice/elevenlabs-agents) | 80.9 | 81.1% | 3.7% | 38.4% | 1.15 | 3.32 |
| 3 | [LiveKit Agents](https://ai.g2.com/evaluations/categories/voice/livekit) | 80.4 | 89.1% | 5.7% | 32.9% | 3.30 | 6.80 |
| 4 | [Pipecat](https://ai.g2.com/evaluations/categories/voice/pipecat) | 77.0 | 86.0% | 11.4% | 40.2% | 3.55 | 6.50 |
| 5 | [Vapi](https://ai.g2.com/evaluations/categories/voice/vapi) | 65.0 | 63.8% | 6.9% | 48.0% | 1.90 | 3.20 |
| 6 | [Bland](https://ai.g2.com/evaluations/categories/voice/bland) | 47.1 | 43.3% | 4.8% | 37.9% | 2.25 | 5.01 |

## Industry results

Healthcare, Legal, and Customer support have separate task-success rankings and passed-call counts in the JSON and CSV exports. Overall scores, WER, and latency are not industry-specific measurements.

## Evidence and interpretation

- Values reproduce the published leaderboard at its displayed precision; they are not recalculated from rounded components or hidden sorting values.
- Platform score, word error rates, and latency are overall measurements across all industries. Industry results report only task success and call counts; no industry-specific WER or latency is supplied.
- The leaderboard uses 80/10/10 task-success/WER/latency weights. The source methodology metric formulas and PDF use 50/25/25. The published values are preserved without reconciling that discrepancy.
- Source scorecard citations link to an older LiveKit-versus-PolyAI study and must not be treated as provider-specific evidence for every current leaderboard score.
- Cite the canonical provider page for a platform's results and the methodology for how measurements were produced. Retain displayed precision; do not derive new scores from rounded components.