Working paper / 2026

MISTT

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.

Leaderboard ↗Read the white paper

v0 · October 2026

← All models

Velma-2 Streaming English

Modulate · Streaming · #20 of 36 streaming by English WER · modulate-velma-2-stt-streaming-english-v2-stream

English WER

4.96%

95% CI 4.63% to 5.30%

TTFS p50 / p95

287 ms

p95 569 ms

$ / 1k audio min

$0.90

list $0.000833/min

Δ Severe

+6.0 pts

3.97% → 10.02% on 100 clips

Accuracy by dataset

English index	4.96%	4.63% – 5.30%
pipecat (voice-agent turns)	3.00%	2.67% – 3.32%
primock57 (doctor–patient)	6.91%	6.32% – 7.47%
Multilingual (de/fr/es/it pooled)	—
VoxPopuli English	6.04%	5.06% – 6.91%

English substitutions 2.43%, deletions 1.59%, insertions 0.94% · raw (un-normalized) English WER 18.27% · hallucination 0.41% of clips · empty 0.14% · VoxPopuli CER 3.60%

Streaming latency

	p50	p90	p95	p99	max
TTFS	287 ms	479 ms	569 ms	775 ms	1.06 s
TTFT	1.25 s	1.59 s	1.68 s	1.83 s	3.54 s
TTFB	1.63 s	1.79 s	1.83 s	1.90 s	2.00 s
Connection setup	201 ms	211 ms	214 ms	220 ms	265 ms

English WER@500 ms 5.09% (final 4.96%) · finals before t0 0.0% · fallback finals 0.0% · partial words later rewritten 11.7%

Robustness · same 100 clips

Mix	WER	Δ pts (95% CI)	Worse	Rank
Clean	3.97%	—	—	#58
Light	3.77%	−0.2 (−0.8 to +0.4)	13/100	#42
Moderate	8.53%	+4.6 (+2.7 to +7.0)	38/100	#44
Severe	10.02%	+6.0 (+3.8 to +8.8)	44/100	#45
Crosstalk	12.90%	+8.9 (+6.7 to +11.3)	54/100	#38

Bleed 3.60% of background-talker words · ranks are out of 81 on the subset.

Configuration and pricing

Languages: en
Finalize: empty text frame = end of audio; ack = done (no flush that keeps the stream open)
Config notes: endpointing=false: one final utterance for the clip, released by the end-of-audio frame.
Rate: $0.000833/min pay-as-you-go · billed on audio time
Effective: $0.90 per 1k audio min on this run · clean run cost $0.25
Pricing notes: Credit-based prepaid, priced per audio hour. Partials only every ~1.5 s by design.
Source: https://platform.modulate.ai/pricing · checked 2026-10-19
Clip outcomes: unsupported 799 · ok 2,200