Environments
Task worlds with tools, state, and verifiable rewards.
Audio data research lab
Advancing AI
You Can Talk To
RL environments for your speech-to-speech models.
Research
Three pieces of the loop
Personalities
Speaker distributions spanning languages, accents, turn-taking behaviour, and acoustic conditions.
Evaluation
Rewards computed from the final state and every tool call, not model judgment.
Releases
Published Work
View all benchmarksBenchmark
Speech-to-speech models on live calls
Task success and reliability of realtime voice models on live phone calls.
PreviewReliability
01GPT Realtime 2.179.3% pass³
02GPT-Live 1 + Sol low75.6% pass³
03Azure Realtime70.7% pass³
Benchmark
Voice agent workflows
End-to-end voice agent platforms on 82 multi-step scenarios.
PreviewReliability
01Retell75.6% pass³
02LiveKit70.7% pass³
03ElevenLabs69.5% pass³
Benchmark
Speech-to-text
Transcription accuracy and latency across 17 STT models.
PreviewWord error rate
01AssemblyAI Universal 3.6 Pro1.77% WER
02AssemblyAI Universal 3.5 Pro1.93% WER
03Google Chirp 32.00% WER
Experiment
Voice quality of production agents
How deployed voice agents sound across 270 live calls.
PreviewScore
01Bland92.8 / 100
02Retell92.2 / 100
03Cresta90.1 / 100
Work with us
Ready to put your voice model to the test?
See how it handles live callers before it matters.