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
StepFun · Streaming · #18 of 36 streaming by English WER · stepfun-stepaudio-2-5-asr-stream-stream
English WER
4.81%
95% CI 4.40% to 5.23%
TTFS p50 / p95
276 ms
p95 438 ms
$ / 1k audio min
$3.25
list $0.003/min
Δ Severe
+11.6 pts
3.08% → 14.66% on 100 clips
| English index | 4.81% | 4.40% – 5.23% |
| pipecat (voice-agent turns) | 2.78% | 2.34% – 3.31% |
| primock57 (doctor–patient) | 6.84% | 6.17% – 7.51% |
| Multilingual (de/fr/es/it pooled) | 9.90% | 8.60% – 11.46% |
| VoxPopuli English | 8.07% | 7.04% – 9.14% |
| VoxPopuli French | 11.05% | 9.27% – 13.10% |
| VoxPopuli Spanish | 8.91% | 7.06% – 11.13% |
English substitutions 2.36%, deletions 1.64%, insertions 0.80% · raw (un-normalized) English WER 20.52% · hallucination 1.16% of clips · empty 0.04% · VoxPopuli CER 6.27%
| p50 | p90 | p95 | p99 | max | |
|---|---|---|---|---|---|
| TTFS | 276 ms | 413 ms | 438 ms | 487 ms | 1.52 s |
| TTFT | 2.85 s | 4.09 s | 4.59 s | 5.66 s | 9.15 s |
| TTFB | 112 ms | 2.11 s | 2.12 s | 2.21 s | 2.46 s |
| Connection setup | 423 ms | 458 ms | 491 ms | 714 ms | 2.63 s |
English WER@500 ms 5.19% (final 4.81%) · finals before t0 0.0% · fallback finals 0.0% · partial words later rewritten 2.3%
| Mix | WER | Δ pts (95% CI) | Worse | Rank |
|---|---|---|---|---|
| Clean | 3.08% | — | — | #40 |
| Light | 4.57% | +1.5 (+0.7 to +2.4) | 19/100 | #59 |
| Moderate | 13.78% | +10.7 (+6.5 to +15.2) | 51/100 | #71 |
| Severe | 14.66% | +11.6 (+7.2 to +16.3) | 49/100 | #70 |
| Crosstalk | 15.35% | +12.3 (+8.3 to +16.3) | 60/100 | #44 |
Bleed 2.71% of background-talker words · ranks are out of 81 on the subset.
