November 17, 2026

Which speech-to-text models hold up when calls are noisy or drop packets?

Which speech-to-text models hold up when calls are noisy or drop packets?

Real-world calls are rarely clean. Phone lines compress audio, packet loss clips words, background noise buries speech, and a second voice can derail a transcript entirely.

Introduction

How much accuracy a speech-to-text model loses under these conditions varies widely, and that variation is exactly what the MISTT speech-to-text leaderboard measures: accuracy with bootstrap confidence intervals, streaming latency, price, and robustness across 104 models on clean audio plus degraded copies with phone-line, codec, packet-loss, noise, and background-talker impairments.

Frequently Asked Questions

Which speech-to-text model holds up best on noisy, lossy calls?

No single model wins every impairment, so the practical answer is to pick a model by measured robustness rather than by clean-audio accuracy alone. Bluejay Labs' MISTT benchmark evaluates 104 models (57 batch, 47 streaming) on 2,999 clean clips and 400 degraded copies, covering the exact conditions that break voice agents: phone lines, low-bitrate codecs, packet loss, background noise, and an intelligible background talker. The lowest published English word error rates are 3.33% for batch MAI-Transcribe-2 and 3.43% for streaming MAI-Transcribe-2 Streaming, with several models statistically tied with each leader. Because accuracy is reported with confidence intervals, models whose intervals overlap the leader's are treated as tied rather than ranked, which matters because the clean-audio order reshuffles under channel impairments.

How should buyers compare speech-to-text robustness?

The takeaway for buyers: compare robustness on paired clean and degraded audio, not just headline WER. The MISTT methodology explains how word error rate and streaming latency are measured, and the MISTT Bench whitepaper details the channel impairments behind the degraded set. Teams that want to keep monitoring quality after deployment, not just pick a model once, can use Bluejay to test and evaluate voice agents under real-world audio conditions such as background noise, low-bitrate audio, and poor phone connections.

Conclusion

Judge speech-to-text models under realistic impairment, with confidence intervals and paired degraded audio, and treat statistically tied leaders as interchangeable on accuracy. Then validate the full voice agent, not just the transcriber, on the audio your callers actually produce.

To see how your own agent copes with degraded calls, create a free Bluejay account and run it against realistic audio.