Working paper / 2026
Multi-condition Impairment Speech-to-Text Bench
Abstract
Multi-condition Impairment Speech-to-Text Bench (MISTT) measures the speech-to-text models that voice agents depend on. Every model hears the same 2,999 frozen clips of voice-agent calls, doctor–patient consultations and five-language parliamentary speech, plus 400 degraded copies of 100 of them: phone lines, low-bitrate codecs, packet loss, background noise and an intelligible background talker. Streaming models are timed the way an agent waits on them, from the forced end of speech to the final transcript. Accuracy carries bootstrap confidence intervals, and the degraded copies are paired with their clean originals, so robustness is measured on the same words.
v0 · October 2026
Amazon · Streaming · #1 of 36 streaming by English WER · amazon-transcribe-stream
English WER
3.81%
95% CI 3.46% to 4.13%
TTFS p50 / p95
555 ms
p95 709 ms · fallback (no flush)
$ / 1k audio min
$11.4
list $0.010002/min
Δ Severe
+6.4 pts
2.40% → 8.77% on 100 clips
| English index | 3.81% | 3.46% – 4.13% |
| pipecat (voice-agent turns) | 2.40% | 2.15% – 2.67% |
| primock57 (doctor–patient) | 5.22% | 4.66% – 5.84% |
| Multilingual (de/fr/es/it pooled) | 12.39% | 11.44% – 13.44% |
| VoxPopuli English | 5.79% | 4.90% – 6.68% |
| VoxPopuli German | 9.60% | 8.33% – 10.88% |
| VoxPopuli French | 10.46% | 8.72% – 12.30% |
| VoxPopuli Spanish | 7.84% | 6.44% – 9.56% |
| VoxPopuli Italian | 20.78% | 18.27% – 23.35% |
English substitutions 1.90%, deletions 0.96%, insertions 0.95% · raw (un-normalized) English WER 16.42% · hallucination 1.44% of clips · empty 0.00% · VoxPopuli CER 8.24%
| p50 | p90 | p95 | p99 | max | |
|---|---|---|---|---|---|
| TTFS | 555 ms | 671 ms | 709 ms | 762 ms | 845 ms |
| TTFT | 579 ms | 656 ms | 721 ms | 1.10 s | 6.14 s |
| TTFB | 185 ms | 639 ms | 647 ms | 658 ms | 1.17 s |
| Connection setup | 8 ms | 10 ms | 11 ms | 14 ms | 115 ms |
English WER@500 ms 3.79% (final 3.81%) · finals before t0 0.3% · fallback finals 100.0% · partial words later rewritten 7.7%
| Mix | WER | Δ pts (95% CI) | Worse | Rank |
|---|---|---|---|---|
| Clean | 2.40% | — | — | #11 |
| Light | 2.93% | +0.5 (+0.2 to +0.9) | 16/100 | #19 |
| Moderate | 9.78% | +7.4 (+5.5 to +9.2) | 63/100 | #52 |
| Severe | 8.77% | +6.4 (+4.1 to +8.9) | 48/100 | #33 |
| Crosstalk | 6.65% | +4.2 (+2.7 to +5.8) | 38/100 | #10 |
Bleed 2.24% of background-talker words · ranks are out of 81 on the subset.
