Papers Accepted at Interspeech 2026!
Two of our papers have been accepted at INTERSPEECH 2026, which will be held in Sydney, Australia. The accepted papers focus on multimodal consistency across speech and text, and on real-world Arabic spoken interactions with LLMs.
🗣️ WASIL: In-the-Wild Arabic Spoken Interactions with LLMs
Title: WASIL: In-the-Wild Arabic Spoken Interactions with LLMs
Authors: Zien Sheikh Ali, Hamdy Mubarak, Soon-Gyo Jung, Hunzalah Hassan Bhatti, Firoj Alam, Shammur Absar Chowdhury
Summary: WASIL is, to our knowledge, the first in-the-wild dataset of Arabic spoken interactions with LLMs. It includes spoken prompts from 93 users across Algeria, Egypt, Sudan, and Syria, covering MSA, multiple dialects, code-switching, and English. The dataset supports analysis of ASR-to-LLM error propagation, direct audio understanding, dialect robustness, and user dissatisfaction patterns in spoken assistant settings.
Session: Wednesday, 30 September, 16:30-18:30
Track: Long Oral 1
Data: https://huggingface.co/datasets/QCRI/WASIL
Paper: https://arxiv.org/abs/2605.16364
🎙️ Said Aloud, Read Different: Cross-Modal Instability in Multimodal Models
Title: Said Aloud, Read Different: Cross-Modal Instability in Multimodal Models
Authors: Basel Mousi, Fahim Dalvi, Shammur Chowdhury, Firoj Alam, Nadir Durrani
Summary: This paper studies whether multimodal models behave consistently when the same input is presented as text or speech. We introduce M2CQA-S, a culturally grounded benchmark with 10,150 image sets from 18 MENA countries, and propose Contrastive Instability (CI) to measure inconsistent model behavior across related statements. The results show that speech input increases instability, especially in Arabic, and that background noise further reduces reliability.
Session: Wednesday, 30 September, 09:00-11:00
Poster: Area 12, Poster 2
Data: https://huggingface.co/datasets/QCRI/M2CQA-S
Code: https://github.com/baselmousi/cfhr-ci
📍 Venue: INTERSPEECH 2026
🌎 Location: Sydney, Australia