2Nd MLC-SLM Challenge Launches, Advancing Multilingual Conversational Speech Understanding

(MENAFN- EIN Presswire) EINPresswire/ -- The 2nd Multilingual Conversational Speech Language Model (MLC-SLM) Challenge has officially opened for registration, inviting research teams and practitioners worldwide to participate. Built on a multilingual conversational speech training set covering 14 languages and approximately 2,100 hours of data, this year's challenge focuses on key tasks including speaker segmentation, automatic speech recognition (ASR), and dialogue understanding, further pushing speech language model research from simple transcription toward deeper conversational understanding.

Targeting Real-World Multilingual Conversations

As speech language models continue to evolve, real-world multilingual conversations are becoming an increasingly important research direction. Unlike conventional ASR tasks, these scenarios involve multiple speakers, multi-turn interactions, and more complex acoustic and semantic information. Systems are expected not only to transcribe speech accurately, but also to determine who spoke when and ultimately understand the conversation as a whole.

The 2nd MLC-SLM Challenge is designed around this shift, focusing on multilingual conversational speech tasks that are closer to real application settings and providing an open benchmark and international platform for Speech LLM research.

Expanded Training Data: Around 2,100 Hours Across 14 Languages

One of the most significant highlights of this year's challenge is the dataset. The training set contains approximately 2,100 hours of multilingual conversational speech spanning 14 languages: English, French, German, Italian, Portuguese, Spanish, Japanese, Korean, Russian, Thai, Vietnamese, Tagalog, Urdu, and Turkish.

Among them, English contributes around 500 hours and includes diverse regional varieties such as US, UK, Australian, Indian, and Philippine English, while each of the other languages contributes roughly 100 hours. This expansion strengthens the challenge's foundation for multilingual conversational speech research in terms of scale, language coverage, and regional diversity.

Natural Two-Speaker Conversations Collected in Realistic Settings

The dataset is designed to better reflect real application scenarios. All recordings are natural two-speaker conversations, where participants discuss randomly assigned topics in a meaningful and fluent way. The audio was collected in quiet indoor environments using consumer devices such as iPhones, making the data closer to real-world collection conditions.

The dataset also includes real-time timestamps and speaker labels to support system development. In addition, Track 1 and Track 2 share the same training set, encouraging participants to explore unified modeling approaches across recognition, diarization, and conversational understanding.

Two Core Tasks: From“Who Spoke” to“What Was Understood”

The challenge includes two main tasks.
Track 1: Multilingual Conversational Speech Diarization and Recognition
Track 2:Multilingual Conversational Speech Understanding

Unlike traditional speech benchmarks that focus primarily on transcription, the 2nd MLC-SLM Challenge places greater emphasis on multilingual, multi-speaker, and dialogue-level understanding. The evaluation setting does not provide prior information such as pre-segmented utterances or speaker labels, making the tasks closer to real deployment conditions.

Building on the International Impact of the First Edition

The new edition builds on the success of the inaugural MLC-SLM Challenge, which was held as a satellite event of Interspeech 2025. The first challenge attracted 78 teams from 13 countries and regions, generated 489 valid leaderboard submissions across two tracks, and received 14 high-quality technical reports. Its summary paper has also been accepted by ICASSP 2026, further demonstrating the challenge's academic value and growing international visibility.

Registration for the 2nd MLC-SLM Challenge is now open

● March 30, 2026: Registration opens
● April 10, 2026: Training data release
● April 24, 2026: Development set and baseline system release
● June 15, 2026: Evaluation set release and leaderboard open
● June 25, 2026: Leaderboard freeze and paper submission portal opens (CMT system)
● July 10, 2026: Paper submission deadline
● July 20, 2026: Notification of acceptance
● October 2, 2026: Workshop date

By offering open data, realistic tasks, and an international exchange platform, the challenge aims to bring together more research teams to advance multilingual conversational speech language modeling. The launch of the second edition also provides a new benchmark for pushing speech language models from simply“hearing clearly” toward genuinely“understanding” conversations.

Registration Links:
Official Website:

MENAFN13042026003118003196ID1110975920