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

Chirp 3 Streaming

Google Cloud · Streaming · #22 of 36 streaming by English WER · google-chirp-3-stream

English WER

5.10%

95% CI 4.68% to 5.50%

TTFS p50 / p95

731 ms

p95 1.03 s

$ / 1k audio min

$18.3

list $0.016/min

Δ Severe

+9.5 pts

3.33% → 12.78% on 100 clips

Accuracy by dataset

English index	5.10%	4.68% – 5.50%
pipecat (voice-agent turns)	1.96%	1.73% – 2.20%
primock57 (doctor–patient)	8.23%	7.48% – 8.99%
Multilingual (de/fr/es/it pooled)	13.13%	12.12% – 14.17%
VoxPopuli English	7.34%	6.27% – 8.34%
VoxPopuli German	13.09%	11.62% – 14.62%
VoxPopuli French	11.70%	9.68% – 13.78%
VoxPopuli Spanish	8.40%	6.85% – 10.14%
VoxPopuli Italian	19.52%	16.84% – 22.12%

English substitutions 2.10%, deletions 2.19%, insertions 0.81% · raw (un-normalized) English WER 18.75% · hallucination 2.74% of clips · empty 0.03% · VoxPopuli CER 8.91%

Streaming latency

	p50	p90	p95	p99	max
TTFS	731 ms	934 ms	1.03 s	1.29 s	3.13 s
TTFT	5.47 s	5.74 s	5.96 s	6.63 s	11.57 s
TTFB	5.73 s	6.59 s	6.95 s	8.64 s	14.86 s
Connection setup	1 ms	1 ms	2 ms	2 ms	145 ms

English WER@500 ms 54.00% (final 5.10%) · finals before t0 0.0% · fallback finals 0.0% · partial words later rewritten 10.8%

Robustness · same 100 clips

Mix	WER	Δ pts (95% CI)	Worse	Rank
Clean	3.33%	—	—	#43
Light	3.81%	+0.5 (−0.8 to +2.3)	11/100	#44
Moderate	11.78%	+8.5 (+5.5 to +11.8)	45/100	#64
Severe	12.78%	+9.5 (+5.5 to +13.8)	44/100	#63
Crosstalk	15.14%	+11.8 (+7.9 to +16.1)	59/100	#43

Bleed 0.93% of background-talker words · ranks are out of 81 on the subset.

Configuration and pricing

Languages: en, de, fr, es, it
Finalize: no flush message: finalize half-closes the request stream; ack = call completes
Config notes: Google suggests ~100 ms chunks; 20 ms keeps parity with the other stream models.
Rate: $0.016/min pay-as-you-go · billed on audio time · 1 s increments
Effective: $18.3 per 1k audio min on this run · clean run cost $7.75
Pricing notes: Pricing page lists 'chirp' under V2 Standard but does not name chirp_3; assumed same Standard Recognition SKU. Streaming not priced separately.
Source: https://cloud.google.com/speech-to-text/pricing · checked 2026-10-19
Clip outcomes: ok 2,999