MCIF-ST: Context-aware Speech Recognition and Speech Translation from MCIF

MCIF-ST provides both long-form and short-form ready-to-use Automatic Speech Recogniton (ASR) and Speech Translation (ST) data derived from MCIF (Multimodal Crosslingual Instruction Following), a multilingual benchmark based on scientific talks. While the original MCIF release packages its content as instruction-following rows (multimodal context + prompt + expected answer, for QA/Summarization/ASR/Translation tasks), MCIF-ST re-exposes the same underlying talk- and sentence-level alignments as plain (audio, source transcript, reference translations) rows -- no instruction wrapper -- for directly evaluating ASR/ST systems, the way you'd use LibriSpeech or CoVoST2.

Two configs are provided over the same 919 sentence-level segments across 21 talks (about 2 hours), differing only in how audio (and, for ST_long, which also contain the paper pdf, useful for context-aware evaluation) is packaged:

• ST_short: audio is physically cut into one clip per segment.
• ST_long: audio is kept as the full, unsegmented talk recording, with offset/duration locating each segment, plus the talk's source pdf.

Dataset Structure (ST_short)

Field	Type	Description
id	string	{talk_id}_{segment_index:04d}, unique per row.
talk_id	string	Talk identifier.
audio	string	Path to the segment's own clip, MCIF_DATA/SHORT_AUDIO/{id}.wav.
offset	float64	Segment start time (s) within the original full talk audio -- kept for provenance even though the clip is already cut.
duration	float64	Segment duration (s); matches the clip's length exactly.
speaker_id	string	Speaker id within the talk.
source_text_en	string	English source transcript of the segment.
target_text_de	string	German reference translation.
target_text_it	string	Italian reference translation.
target_text_zh	string	Chinese reference translation.

Dataset Structure (ST_long)

Field	Type	Description
id	string	Same scheme as ST_short -- matching ids across the two configs describe the same segment.
talk_id	string	Talk identifier.
audio	string	Path to the full, unsegmented talk recording, MCIF_DATA/LONG_AUDIO/{talk_id}.wav (shared by all segments of that talk).
pdf	string	Path to the talk's slide/paper pdf, MCIF_DATA/PDF/{talk_id}.pdf.
offset	float64	Segment start time (s) within audio.
duration	float64	Segment duration (s).
speaker_id	string	Speaker id within the talk.
source_text_en	string	English source transcript of the segment.
target_text_de	string	German reference translation.
target_text_it	string	Italian reference translation.
target_text_zh	string	Chinese reference translation.

Usage

from datasets import load_dataset

# ST_short: pre-cut clips, load and use directly
ds_short = load_dataset("FBK-MT/MCIF-ST", "ST_short", split="test")
row = ds_short[0]
print(row["source_text_en"], "->", row["target_text_de"])

# ST_long: full audio, slice using offset/duration if you want to do short-form inference
# otherwise, load directly the files contained in MCIF_DATA/LONG_AUDIO (read below)
import soundfile as sf
ds_long = load_dataset("FBK-MT/MCIF-ST", "ST_long", split="test")
row = ds_long[0]
audio, sr = sf.read(row["audio"])
start, end = int(row["offset"] * sr), int((row["offset"] + row["duration"]) * sr)
segment_audio = audio[start:end]

For an easy evaluation with resegmentation tools (e.g., mWERsegmenter), audio_segments.yaml file is provided containing the segments alignment of the long-form audio contained in MCIF_DATA/LONG_AUDIO (with also the corresponding pdf files for each talk in MCIF_DATA/PDF), together with {en,de,it,zh}.txt reference files, with one segment per line.

License

CC-BY-4.0, matching the source MCIF release.

Citation

If you use this data, please cite the original MCIF paper:

@inproceedings{papi2026mcif,
title={{MCIF}: Multimodal Crosslingual Instruction-Following Benchmark from Scientific Talks},
author={Sara Papi and Maike Z{\"u}fle and Marco Gaido and Beatrice Savoldi and Danni Liu and Ioannis Douros and Luisa Bentivogli and Jan Niehues},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=PtPYZYfa0h}
}