[Last update: Mar. 6, 2025]

Description

Motivation

Large language models (LLMs) have demonstrated the capability of performing several NLP tasks without the need for building dedicated models, offering a single solution for many applications. While solely processing text in their initial stage, LLMs are now being enhanced by integrating other modalities, like vision and audio. In this scenario, the emerging paradigm of creating a unique architecture from speech foundation models (SFMs) and LLMs is gaining traction to combine the best of both worlds: the ability to process spoken language inputs with the always-evolving language knowledge of the LLMs. For this reason, this year, we introduce the first shared task at IWSLT aimed at testing general and universal models for the speech modality.

Tasks Description

Participants are asked to build a model capable to perform, depending on the track, the following tasks:

  • SHORT TRACK (input: automatically segmented audio):
    • Automatic Speech Recognition (ASR): the speech is transcribed into the same language;
    • Speech-to-text Translation (S2TT): the speech is translated into the target language;
    • Spoken Question Answering (SQA): textual questions have to be answered based on the spoken content in the same language and in a language different from the speech (questions and answers are always in the same language);
  • LONG TRACK (input: long-form audio):
    • Automatic Speech Recognition (ASR): the speech is transcribed into the same language;
    • Speech-to-text Translation (S2TT): the speech is translated into the target language;
    • Spoken Question Answering (SQA): textual questions have to be answered based on the spoken content in the same language and in a language different from the speech (questions and answers are always in the same language);
    • Speech-to-text Summarization (S2TSUM): a summary has to be provided from the spoken content in the same language and in a language different from the speech. All tasks listed for each track are mandatory.

Languages

English for ASR, monolingual SQA and S2TSUM, and English -> German, Italian, Chinese for S2TT, multilingual SQA and S2TSUM.

IMPORTANT! The results can be submitted for some or all language directions.

Data Conditions

We adopt two conditions. The first is constrained, where a pre-defined training setting is adopted, and only a specified pre-trained SFM and LLM architecture can be used. The second is unconstrained, with no limitation on pre-trained models and training data.

Constrained

Participants are allowed to use the SFM and LLM provided below, and training the system on the following data.

We do not provide any training data for SQA (questions and answers) and S2TSUM (summaries) in languages different from the source speech.

IMPORTANT! The use of the pre-trained SFM and/or LLM is NOT mandatory and we also accept submissions with models trained from scratch on the allowed data as well as solutions using only one of the two pre-trained models (either SFM or LLM).

Unconstrained

Any model, any data.

Evaluation

We release the video, the source audio, and the instructions, and participants submit their outputs. The instructions can be modified by participants to match their system’s prompts. In SQA, questions are provided both in the same language of the speech (English) and in different languages (German, Italian, Chinese) but they always have to be replied to in the same language of the questions (e.g., an Italian question should be replied to in Italian). Questions can also be nonanswerable, in this case, only the answer “Not answerable.” (and the corresponding Italian “Non è possibile rispondere.”, German “Nicht zu beantworten.”, and Chinese “无法回答。” translations) will be considered correct.

The Long Track will process audio files in WAV format that are, on average, 5–6 minutes long. The Short Track will handle the same audio files, but they will be automatically segmented into 15–20 second audio segments, on average, using SHAS.

We provide an example for the Long track, downloadable here. Participants are also allowed to use it as 1-shot example for their model.

We also provide useful scripts for parsing inputs and outputs, downloadble here.

Details on the metrics used for the final ranking will be given after the Evaluation period.

Submission

More details will be given soon.

Organizers

  • Sara Papi, Fondazione Bruno Kessler
  • Luisa Bentivogli, Fondazione Bruno Kessler
  • Marco Gaido, Fondazione Bruno Kessler
  • Danni Liu, Karlsruhe Institute of Technology
  • Beatrice Savoldi, Fondazione Bruno Kessler
  • Maike Züfle, Karlsruhe Institute of Technology

Contact

Chair(s): Sara Papi (spapi@fbk.eu)

Discussion: iwslt-evaluation-campaign@googlegroups.com