Multilingual TEDx (mTEDx) — SLR100

Dataset Description

Homepage: OpenSLR SLR100

mTEDx is a multilingual speech recognition and translation corpus built from TEDx Talks.
Original resource: https://www.openslr.org/100/

The corpus provides audio recordings and VTT transcripts for 8 languages (Spanish, French, Portuguese, Italian, Russian, Greek, Arabic, German) with aligned translations into up to 5 languages (English, Spanish, French, Portuguese, Italian).

License: CC BY-NC-ND 4.0
Contact: Elizabeth Salesky (esalesky@jhu.edu), Matthew Wiesner (wiesner@jhu.edu)

---

Corpus Statistics

Each row in the dataset corresponds to one segment (individual audio clip + transcript).
The table below reflects sentence counts and total audio duration as reported in docs/statistics.txt per language.

Arabic (ar)

Split	Talks	Sentences	Words	Duration
train	95	11 821	115 259	68 310 s ≈ 18h 58m
valid	7	1 079	9 374	5 280 s ≈ 1h 28m
test	7	1 066	8 964	5 187 s ≈ 1h 26m
total	109	13 966	133 597	78 778 s ≈ 21h 53m

German (de)

Split	Talks	Sentences	Words	Duration
train	53	6 764	94 984	44 958 s ≈ 12h 29m18s
valid	9	1 172	14 661	6 893 s ≈ 1h 54m53s
test	9	1 166	14 289	6 715 s ≈ 1h 51m55s
total	71	9 062	123 934	58 566 s ≈ 16h 16m6s

All Languages — Download Sizes (original tarballs)

Config	Language	Tarball size
es	Spanish	35 GB
fr	French	34 GB
pt	Portuguese	29 GB
it	Italian	19 GB
ru	Russian	10 GB
el	Greek	5.5 GB
ar	Arabic	3.6 GB
de	German	2.6 GB

---

Dataset Structure

Schema

Each example corresponds to one audio segment extracted from a full TEDx talk using the Kaldi segments timestamps file.

Field	Type	Description
id	string	Unique segment id: <talk_stem>_<index> (e.g. 14zpc3Nj_e4_0003)
audio	Audio	Audio float32 waveform of the segment
transcript	string	Transcription text
duration	float32	Duration of the audio segment in seconds
talk_id	string	Source talk file stem
segment_id	int32	0-based index of the segment within its talk
start	float32	Segment start time within the source talk (seconds)
end	float32	Segment end time within the source talk (seconds)

Splits

Split	Description
train	Training set
valid	Validation / development set
test	Test set

---

Usage

from datasets import load_dataset

# Load Arabic training split
ds = load_dataset("deepdml/mtedx", "ar", split="train")
print(ds[0])
# {
#   'id':         '14zpc3Nj_e4_0001',
#   'audio':      {'array': array([...], dtype=float32), 'sampling_rate': 16000},
#   'transcript': 'أكل العالم وغص بنخلة',
#   'duration':   4.16,
#   'talk_id':    '14zpc3Nj_e4',
#   'segment_id': 1,
#   'start':      9.332,
#   'end':        13.492,
#   'language':   'ar'
# }

# Stream a large language without downloading everything
ds = load_dataset("deepdml/mtedx", "es", split="train", streaming=True)
for sample in ds:
    audio = sample["audio"]["array"]           # numpy float32 array @ 16 kHz
    text  = sample["transcript"]
    dur   = sample["duration"]                 # seconds
    break

# ASR fine-tuning example (Whisper / wav2vec2)
ds = load_dataset("deepdml/mtedx", "fr", split="train")
ds = ds.select_columns(["audio", "transcript", "duration"])

---

Source Data

Downloaded from OpenSLR SLR100.
Each language pack (mtedx_<lang>.tgz) contains:

• data/<split>/wav/ — Full-talk FLAC audio files
• data/<split>/vtt/ — WebVTT transcript files (<id>.<lang>.vtt)
• data/<split>/txt/ — Segments and Plain-text transcripts
• docs/statistics.txt — Per-split statistics

The upload script (create_mtedx_dataset.py) slices the full-talk FLAC files into individual segments using the kaldi segments timestamps and discards segments shorter than 0.5 s or longer than 30 s.

---

Citation

@inproceedings{salesky2021mtedx,
  title     = {Multilingual TEDx Corpus for Speech Recognition and Translation},
  author    = {Elizabeth Salesky and Matthew Wiesner and Jacob Bremerman and
               Roldano Cattoni and Matteo Negri and Marco Turchi and
               Douglas W. Oard and Matt Post},
  booktitle = {Proceedings of Interspeech},
  year      = {2021},
}