Datasets:

maikezu
/

iwslt2026-metrics-shared-train-dev

dataset_info:
  features:
    - name: audio_path
      dtype: string
    - name: doc_id
      dtype: string
    - name: src_text
      dtype: string
    - name: src_text_system
      dtype: string
    - name: src_lang
      dtype: string
    - name: tgt_lang
      dtype: string
    - name: domain
      dtype: string
    - name: tgt_system
      dtype: string
    - name: tgt_text
      dtype: string
    - name: score
      dtype: float64
    - name: audio
      dtype: audio
  splits:
    - name: train
      num_bytes: 8046291442
      num_examples: 33721
    - name: train_synthetic
      num_bytes: 324856288
      num_examples: 7000
    - name: dev
      num_bytes: 1241951905
      num_examples: 5556
  download_size: 71228386133
  dataset_size: 9613099635
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: train_synthetic
        path: data/train_synthetic-*
      - split: dev
        path: data/dev-*
license: cc-by-nc-nd-4.0
tags:
  - quality-estimation
  - speech-translation
  - speech
size_categories:
  - 10K<n<100K

Metrics-Shared Task IWSLT 2026 Train and Dev Set

This dataset contains the train and dev sets for the Speech Translation Metrics Shared Task at IWSLT 2026. More details about the shared task can be found on the IWSLT website.

The dataset is primarily designed for research in speech translation quality estimation.

Task Goal

Given a speech sample and a system-generated translation, the goal is to estimate a score that reflects the translation quality.

Dataset Splits

train

Contains a mix of:

IWSLT 2023 human annotations (details IWSLT 2023)
- Previous and following segments can be inferred from the doc_id feature.
- Human evaluators considered context of one previous and one following segment.
WMT 2024 human annotations (details WMT 2024)
- Evaluated on segmented audio. The information on previous/following segments is not available.
WMT 2025 human annotations (details WMT 2025)
- Evaluated on segmented audio. The information on previous/following segments is not available.

train_synthetic

Contains:

SpeechQE data (details SpeechQE)
- Based on Common Voice.
- Automatically annotated (synthetic scores).

dev

Contains:

IWSLT 2025 ACL Talks human annotations (details IWSLT 2025)
- Previous and following segments can be inferred from the doc_id feature.
- Human evaluators considered context of one previous and one following segment.

Features

Column	Type	Description
`audio`	`Audio`	The speech waveform of the segment
`audio_path`	`string`	Path to the audio file
`doc_id`	`string`	Unique identifier for the segment/document
`src_text`	`string`	Source text
`src_text_system`	`string`	Source text system (e.g. human, ASR model)
`src_lang`	`string`	Source language code (e.g., `en`)
`tgt_text`	`string`	Target text translation
`tgt_lang`	`string`	Target language code (e.g., `de`)
`domain`	`string`	Domain or dataset source
`tgt_system`	`string`	Target system or model used for translation
`score`	`float64`	Human or synthetic evaluation score (0–1)

Dataset Card Contact

Maike Züfle @maikezu