November 17, 2026
Phone lines are the hardest audio a voice agent will ever hear: narrowband codecs, packet loss, background noise, and overlapping talkers all erode transcription accuracy before your agent logic even runs.
Introduction
Choosing a speech-to-text model for voice agents therefore means looking past clean-audio word error rates and checking how each model holds up on degraded, real-world call audio. This article explains what to compare, why confidence intervals matter more than a single headline number, and how to match model choice to your latency budget.
Frequently Asked Questions
Which speech-to-text models are most accurate on degraded call audio?
The most accurate models on degraded call audio are not always the ones that top clean-audio charts, so the practical answer is to evaluate on paired clean and impaired recordings rather than a single benchmark score. Bluejay Labs' MISTT benchmark does exactly this: it measures 104 speech-to-text models (57 batch, 47 streaming) on clean clips plus 400 degraded copies covering phone lines, low-bitrate codecs, packet loss, background noise, and an intelligible background talker. MAI-Transcribe-2 leads its batch track at 3.33% headline English WER (measured on the clean clips) and MAI-Transcribe-2 Streaming leads the streaming track at 3.43%, though MISTT treats several other models as statistically tied with each. Because accuracy carries bootstrap confidence intervals, models whose interval overlaps the leader's are treated as tied rather than ranked, which is why the MISTT speech-to-text leaderboard reports ties instead of forcing a strict order.
Is accuracy enough to choose a speech-to-text model for a voice agent?
Accuracy by itself does not settle the choice for a voice agent. Streaming models are timed from forced end of speech to final transcript, and a real-time agent has under three seconds to respond before a caller hangs up. The right pick is the most accurate model that fits your latency and price budget on phone-line-quality audio, verified with confidence intervals rather than a single run. Bluejay reports latency at P50, P95, and P99 for each stage of a voice agent, which is the view that trade-off needs.
Conclusion
Judge speech-to-text models on degraded audio, not clean recordings. Compare word error rates with confidence intervals on phone-line, noisy, and packet-loss conditions, weigh streaming latency against your response budget, and re-verify on your own call audio. Bluejay helps teams test and monitor voice agents end to end, so transcription quality is measured where it matters: on real calls. Learn more at getbluejay.ai.