urgent-challenge/urgent2026-sqa

URGENT 2026 Speech Quality Assessment

The subjective listening-test dataset from ICASSP 2026 URGENT Track 1: Universal Speech Enhancement, containing 5,040 ACR samples, 12,600 CCR comparisons, and 141,386 individual ratings across 840 utterances, nine languages, and six systems.

Each scored sample includes the audio, its MOS or CMOS, individual listener scores, and utterance metadata.

Load the data

from datasets import load_dataset

repo = "urgent-challenge/urgent2026-sqa"
acr = load_dataset(repo, "acr", split="test")
ccr = load_dataset(repo, "ccr", split="test")

sample = acr[0]
print(sample["sample_id"], sample["mos"], sample["listener_scores"])
audio = sample["audio"].get_all_samples()
waveform, sample_rate = audio.data, audio.sample_rate

Both configurations contain a single test split covering the complete ICASSP blind listening test.

Configuration	Contents	Samples
acr	Enhanced audio, MOS, and individual quality scores	5,040
ccr	Enhanced audio pairs, CMOS, and individual comparison scores	12,600

Audio is stored as embedded mono FLAC at its native sample rate. Audio decoding uses TorchCodec; install compatible PyTorch, TorchCodec, and FFmpeg versions. To read the audio bytes with SoundFile instead:

import io
import soundfile as sf
from datasets import Audio

raw_acr = acr.cast_column("audio", Audio(decode=False))
waveform, sample_rate = sf.read(io.BytesIO(raw_acr[0]["audio"]["bytes"]))

ACR: absolute category rating

Listeners rate the quality of a single enhanced audio sample on a 1–5 scale, with higher scores indicating better quality. Each of the six systems has one sample for every utterance. There are 40,280 individual ratings, with 4–16 scores per sample.

Field	Type	Description
sample_id	string	Unique sample ID: acr_{utterance_id}_{system_id}.
audio	Audio	Enhanced speech from the evaluated system.
utterance_id	string	Source utterance ID, such as fileid_1.
system_id	string	Submission ID of the evaluated system.
team_name	string	Name of the submitting team.
language	string	ISO 639-3 language code.
speaker_id	string	Anonymous speaker grouping key.
text	string	Utterance transcript.
is_simulated	bool	Whether the source utterance was simulated.
sample_rate	int32	Native audio sample rate in Hz.
duration	float64	Audio duration in seconds.
mos	float64	Mean opinion score, rounded to three decimal places.
listener_scores	list[int8]	Individual quality scores on the 1–5 scale.

CCR: comparison category rating

Listeners compare two systems' outputs for the same utterance on a −3 to +3 scale. Positive CMOS means audio_a is better than audio_b; negative values favor audio_b. All 15 unordered pairs of the six systems are included for each utterance. There are 101,106 individual ratings, with 4–20 scores per pair.

Field	Type	Description
sample_id	string	Unique pair ID: ccr_{utterance_id}_{system_id_a}_{system_id_b}.
audio_a	Audio	Enhanced speech from system A.
audio_b	Audio	Enhanced speech from system B.
utterance_id	string	Source utterance ID shared by both audio signals.
system_id_a	string	Submission ID of system A.
system_id_b	string	Submission ID of system B.
team_name_a	string	Name of the team that submitted system A.
team_name_b	string	Name of the team that submitted system B.
language	string	ISO 639-3 language code.
speaker_id	string	Anonymous speaker grouping key.
text	string	Utterance transcript.
is_simulated	bool	Whether the source utterance was simulated.
sample_rate	int32	Native sample rate of both audio signals in Hz.
duration	float64	Duration of both audio signals in seconds.
cmos	float64	Mean comparison score, rounded to three decimal places. Positive values favor A.
listener_scores	list[int8]	Individual comparison scores on the −3 to +3 scale, following the A/B order.

System A has the lower numeric submission ID. If you swap A and B, negate both cmos and every value in listener_scores.

Systems

Submission ID	Team	Challenge rank
1412	WR	1
1397	GHW	2
1401	subatomicseer	3
1395	baird	4
1348	RICK2000	5
1399	Ali-Universal-SE	5

Languages

Language	Code	Utterances
English	eng	163
Chinese	zho	117
Arabic	ara	80
German	deu	80
French	fra	80
Hindi	hin	80
Italian	ita	80
Japanese	jpn	80
Korean	kor	80

Citation

If you use this dataset, please cite it by referencing the following paper

@inproceedings{urgent2026,
  title={{ICASSP 2026 URGENT Speech Enhancement Challenge}},
  author={Li, Chenda and Wang, Wei and Sach, Marvin and Zhang, Wangyou and Saijo, Kohei and Cornell, Samuele and Fu, Yihui and Ni, Zhaoheng and Fingscheidt, Tim and Watanabe, Shinji and Qian, Yanmin},
  booktitle={Proc. ICASSP 2026},
  year={2026}
}