Description

Simultaneous translation (also known as real-time or streaming translation) is the task of generating translations incrementally given partial input only. Simultaneous translation enables interesting applications such as automatic simultaneous interpretation or international conference translations. Simultaneous systems are typically evaluated with respect to quality and latency.

There will be two tracks:

  • Text-to-Text: simultaneously translating text in source language into text in target language. Participants will employ an ASR agent of their choice to produce the text, so as to respect computational latency constraints.
  • Speech-to-Text: simultaneously translating speech in source language into text in target language.

in the following language directions (more details will be made available soon):

  • English -> German
  • English -> Arabic (new) (cancelled in the simultaneous track)
  • English -> Chinese
  • English -> Japanese
  • Czech -> English

We have two focuses this year:

  • long-form speech: our evaluation will be conducted on unsegmented speech
  • large language models: participants are allowed to use LLMs (details will be announced later)

The test set domains are the subsets of the ones of the offline track:

  • English -> German: ACL talks and accent challenge data
  • English -> Arabic: business news (cancelled)
  • English -> Chinese: ACL talks
  • English -> Japanese: ACL talks
  • Czech -> English:

Data

The data condition for this task is “constrained with large language models (LLMs)”. Any model that is open-weights with a permissive license is acceptable for use. In addition, pretrained speech encoders and ASR models may be employed.

English-to-X

Our English-to-X training data condition follows the one in the offline task. The list is available here. ACL 60/60 dataset can be the development test set. The development data can be found here while the yaml files containing the audio information (useful for metrics computation) can be found here.

Czech-to-English

Allowed data:

  • ParCzech 3.0 (ASR):
    • Allowed data: parczech-3.0-asr-train-20*.tar.gz
    • https://lindat.mff.cuni.cz/repository/xmlui/handle/11234/1-3631?show=full
  • VoxPopuli (ST)
    • Unlabelled Data: cs_v2
    • translated data Cs -> En data
    • speech-to-speech Cs -> En data
    • https://github.com/facebookresearch/voxpopuli
  • Common Voice Corpus 20.0 (ASR)
    • Czech ASR data
    • CV version: 20.0
    • https://commonvoice.mozilla.org/en/datasets
  • Czeng 2.0 (MT)
    • https://ufal.mff.cuni.cz/czeng
  • OpenSubtitles v2018 (MT)
    • https://opus.nlpl.eu/OpenSubtitles/cs&en/v2018/OpenSubtitles
  • Europarl (MT)
    • https://www.statmt.org/europarl/
  • MOSEL (transcripts only)
    • automatic transcripts for unlabeled VoxPopuli audio
    • https://huggingface.co/datasets/FBK-MT/mosel
  • Dev Set (ST)
    • https://drive.google.com/file/d/1-XicsrBQubkGK-kyBIxKO-7JAx94o_KV/view?usp=sharing

Test set

This year’s test sets will be available at the beginning of the evaluation period (April 1st).

Baselines

The baselines for each language pair can be found here (GitHub).

Submission

The evaluation implementation will use the latest SimulEval toolkit. Participants have two options for the submission:

  • Docker image submission; the organizers run the system to compare the computation-aware latency
  • System log submission; the computation-aware latency cannot be compared directly but will be reported with its hardware difference

Systems submitted via docker image are expected to run on a single NVIDIA A100 GPU with 80 GB of HBM. Additionally, participants must include a README with instructions on how to run the system for each track and language direction.

Regardless of the submission type (Docker or log), participants must also submit results on the development set (i.e., ACL 60/60 or the dedicated Czech-to-English dev set) to determine the latency regime of their submission.

Submission link: (available soon)

Participants can update their submissions during the evaluation period.

Evaluation

Metrics

The system’s performance will be evaluated in two ways:

  • Quality:
    • BLEU
    • Additional results using neural metrics (COMET, BLEURT, …)
  • Latency:

For latency measurement, we will contrast computation aware and non computation aware latency metrics.

Ranking

The systems will be ranked by the translation quality within the latency constraints, measured by non-computation-aware StreamLAAL.

This year, we have two latency regimes, low and high.

The detailed latency constraints (non-computationally-aware StreamLAAL) for each language pair are the following:

  • English-to-German and Czech-to-English: 0-2s (low), 2-4s (high);
  • English-to-Chinese: 0-2.5s (low), 2.5-4s (high);
  • English-to-Japanese: 0-3.5s (low), 3.5s-5s (high).

Human Evaluation

Human evaluation will be conducted for primary submissions.

Organizers

  • Victor Agostinelli (Oregon State University)
  • Lizhong Chen (Oregon State University)
  • Sara Papi (FBK)
  • Peter Polák (Charles University)
  • Katsuhito Sudoh (Nara Women’s University)

Contact

Chair(s): Katsuhito Sudoh (Nara Women’s University)

Discussion: iwslt-evaluation-campaign@googlegroups.com