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

GPT Realtime Whisper

OpenAI · Streaming · #27 of 36 streaming by English WER · openai-gpt-realtime-whisper-stream

English WER

5.82%

95% CI 5.28% to 6.41%

TTFS p50 / p95

563 ms

p95 707 ms

$ / 1k audio min

$18.4

list $0.017/min

Δ Severe

+9.3 pts

4.25% → 13.54% on 100 clips

Accuracy by dataset

English index	5.82%	5.28% – 6.41%
pipecat (voice-agent turns)	2.03%	1.80% – 2.28%
primock57 (doctor–patient)	9.61%	8.49% – 10.60%
Multilingual (de/fr/es/it pooled)	13.37%	12.15% – 14.58%
VoxPopuli English	7.69%	6.58% – 8.80%
VoxPopuli German	11.76%	10.45% – 13.17%
VoxPopuli French	12.37%	10.39% – 14.64%
VoxPopuli Spanish	9.46%	7.18% – 12.13%
VoxPopuli Italian	19.52%	16.95% – 22.07%

English substitutions 2.38%, deletions 2.77%, insertions 0.67% · raw (un-normalized) English WER 20.51% · hallucination 2.77% of clips · empty 0.13% · VoxPopuli CER 8.84%

Streaming latency

	p50	p90	p95	p99	max
TTFS	563 ms	672 ms	707 ms	851 ms	3.00 s
TTFT	1.05 s	1.26 s	1.35 s	1.82 s	7.76 s
TTFB	1.41 s	2.19 s	2.52 s	3.92 s	21.50 s
Connection setup	357 ms	394 ms	419 ms	605 ms	5.96 s

English WER@500 ms 5.88% (final 5.82%) · finals before t0 0.0% · fallback finals 0.1% · partial words later rewritten 2.2%

Robustness · same 100 clips

Mix	WER	Δ pts (95% CI)	Worse	Rank
Clean	4.25%	—	—	#65
Light	4.69%	+0.4 (−0.5 to +1.4)	21/100	#62
Moderate	12.74%	+8.5 (+5.7 to +11.2)	53/100	#69
Severe	13.54%	+9.3 (+6.1 to +12.8)	55/100	#66
Crosstalk	16.43%	+12.2 (+9.6 to +14.9)	69/100	#48

Bleed 3.17% of background-talker words · ranks are out of 81 on the subset.

Configuration and pricing

Languages: en, de, fr, es, it
Finalize: input_audio_buffer.commit; ack = next ...input_audio_transcription.completed
Config notes: turn_detection null (model requires it): only our input_audio_buffer.commit at the VAD speech end produces the final.
Rate: $0.017/min pay-as-you-go · billed on audio time
Effective: $18.4 per 1k audio min on this run · clean run cost $7.80
Pricing notes: Priced by realtime audio duration ($0.017/min); partial-minute rounding not published.
Source: https://developers.openai.com/api/docs/pricing · checked 2026-10-19
Clip outcomes: ok 2,999