ICASSP 2027 Signal Processing Grand Challenge
Embodied Speech Perception under Oracle Navigation in Multi-Party Conversations
EAR-1 asks a foundational question: if navigation is executed correctly, to what extent can movement itself improve multi-party speech perception?
Official status: EAR-1 has been accepted as an ICASSP 2027 Signal Processing Grand Challenge.
Recording paradigm
Scientific motivation
Fixed far-field systems are vulnerable to distance, background noise, reverberation, and overlapping speech. Embodied devices can rotate toward active speakers or approach sound sources, potentially improving direct-to-reverberant ratio and spatial separability.
Movement also introduces self-noise and rapidly varying acoustic transfer functions. EAR-1 isolates this trade-off by using a human proxy to execute a predefined oracle-navigation policy, while participating systems receive only multi-channel audio.
Navigation is controlled. Speech perception remains the challenge.
Challenge tracks
Each task is evaluated separately on synchronized Fixed-Array and Moving-Array recordings of the same conversations.
ESD-ON
Assign consistent speaker labels to continuous multi-channel recordings, including regions with overlapping speech. Oracle voice activity detection regions are provided without revealing active-speaker identities.
ESR-ON
Generate speaker-attributed transcripts from continuous, unsegmented multi-channel recordings. Systems jointly handle speech activity detection, speaker attribution, and recognition.
Dataset
The EAR-1 corpus will provide paired multi-channel recordings of real multi-party conversations captured synchronously by fixed and moving microphone arrays. A labeled development set will support model training and system development, while an evaluation set with hidden annotations will be used for final ranking. The two sets will be disjoint in both rooms and speakers.
Evaluation
Fixed and moving conditions use the same protocols and are ranked independently, enabling controlled within-conversation comparison.
Zero forgiveness collar. Overlapping speech is scored. With oracle VAD, the metric mainly reflects speaker confusion and overlap attribution.
Speaker-wise hypotheses are concatenated in time and scored under the minimum-error mapping between reference and hypothesized speakers.
Schedule
All deadlines use Anywhere on Earth unless otherwise announced.
Participation
Public datasets and pretrained models are allowed when cited, accessible to other teams, and fully disclosed. New resources cannot be introduced after evaluation data release.
Systems may use only the officially released audio and task metadata. Trajectories, poses, auxiliary sensors, and organizer-only annotations are prohibited.
Each entry includes outputs in the official format and a report describing architecture, training, external resources, and decoding configuration.
Organizing team
USTC, China
hangchen@ustc.edu.cnUSTC, China
jundu@ustc.edu.cnBrno University of Technology
ihan@fit.vut.czSoochow University, China
zhu.mengyao@suda.edu.cnAISHELL, China
buhui@aishelldata.comAISHELL, China
xuxin@aishelldata.comUSTC, China
wangruoyu@mail.ustc.edu.cnUSTC, China
jhzhao@mail.ustc.edu.cnUSTC, China
spzhang71@mail.ustc.edu.cnUSTC, China
shuhangwu@mail.ustc.edu.cnUniversity of Palermo, Italy
siniscalchi77.19@gmail.comDeepCybo, China
chenkai@deepcybo.comiFLYTEK, China
jiapan@iflytek.comAnhui University, China
tuyanhui@ahu.edu.cnAISPEECH, China
qiang.zhou@aispeech.comUniversity of Science and Technology Beijing, China
qianxy@ustb.edu.cnJoin the challenge
