From the source
Baseten says it sits at the top of Coval's early-access voice AI benchmarks, placing its speech-to-text deployment on the quality-latency Pareto frontier.
- Coval is a voice AI evaluation platform that runs reproducible, independent benchmarks, and its September 2026 early-access results measure what end users experience: speed and accuracy together for speech-to-text, and time to first audio for text-to-speech.
- Baseten says its STT deployment is about 5× faster than
OpenAI's while also showing the best word error rate, which it argues shows optimized inference can improve responsiveness without sacrificing transcription quality.
- Baseten says it defines the frontier with Qwen3 ASR 1.7B Streaming for STT and Qwen3 TTS 12Hz Base Streaming 1.7B for TTS, and that no provider in the current benchmark is faster while also producing a lower error rate, as of Oct 1., 2026.
- It also leads the TTS quality benchmark and sits on the quality-latency frontier there.
- Baseten argues a voice agent's latency compounds across speech-to-text, reasoning, text-to-speech, orchestration and networking, so a strong median latency can hide a p95 that hurts some conversation turns.
- Coval's benchmark runs from a fixed region for fairness and reproducibility, and Baseten says its deployments enter it as normal API endpoints backed by dedicated inference hardware.





