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

Nemotron 3 ASR Streaming 160 ms

NVIDIA · Streaming · #32 of 36 streaming by English WER · nvidia-cache-aware-parakeet-rnnt-en-us-asr-streaming-160ms-stream

English WER

6.94%

95% CI 6.46% to 7.43%

TTFS p50 / p95

185 ms

p95 198 ms

$ / 1k audio min

free trial tier

Δ Severe

+11.1 pts

4.97% → 16.03% on 100 clips

Accuracy by dataset

English index	6.94%	6.46% – 7.43%
pipecat (voice-agent turns)	4.14%	3.74% – 4.53%
primock57 (doctor–patient)	9.74%	8.84% – 10.62%
Multilingual (de/fr/es/it pooled)	—
VoxPopuli English	8.48%	7.47% – 9.47%

English substitutions 3.41%, deletions 2.72%, insertions 0.81% · raw (un-normalized) English WER 21.52% · hallucination 0.41% of clips · empty 0.18% · VoxPopuli CER 4.89%

Streaming latency

	p50	p90	p95	p99	max
TTFS	185 ms	195 ms	198 ms	317 ms	453 ms
TTFT	717 ms	1.16 s	1.38 s	1.92 s	6.38 s
TTFB	317 ms	319 ms	322 ms	468 ms	1.01 s
Connection setup	1 ms	1 ms	1 ms	1 ms	78 ms

English WER@500 ms 6.94% (final 6.94%) · finals before t0 0.0% · fallback finals 0.0% · partial words later rewritten 3.3%

Robustness · same 100 clips

Mix	WER	Δ pts (95% CI)	Worse	Rank
Clean	4.97%	—	—	#74
Light	6.05%	+1.1 (+0.4 to +1.8)	23/100	#76
Moderate	13.02%	+8.1 (+5.4 to +10.8)	59/100	#70
Severe	16.03%	+11.1 (+7.7 to +14.5)	61/100	#71
Crosstalk	18.43%	+13.5 (+10.1 to +17.2)	72/100	#52

Bleed 4.86% of background-talker words · ranks are out of 81 on the subset.

Configuration and pricing

Languages: en
Finalize: no flush message: half-close the request stream; ack = RPC ends OK after the last final
Config notes: Nemotron 3 ASR streaming (ai-nemotron-asr-streaming). The hosted function serves only this one build, whose interim results advance in 160 ms steps; the 80/560/1120 ms builds are not hosted.
Rate: $0/min pay-as-you-go · billed on audio time
Effective: $0 per 1k audio min on this run · clean run cost $0
Pricing notes: build.nvidia.com hosted endpoints are free, rate-limited, prototyping-only (no published per-minute price). Production = self-hosted NIM under NVIDIA AI Enterprise (per-GPU licence, ~$4,500/GPU/yr per third-party guides) plus GPU cost.
Source: https://build.nvidia.com/nvidia/nemotron-asr-streaming · checked 2026-10-19
Clip outcomes: unsupported 799 · ok 2,200