kadirnar's picture
codecbench multilingual v1
6456017 verified
|
Raw History Blame Contribute Delete
2.45 kB
metadata
license: cc-by-4.0
language:
  - en
  - de
  - fr
  - es
  - it
  - pt
  - nl
  - pl
  - cs
  - tr
  - id
  - vi
  - ru
  - uk
  - el
  - ar
  - fa
  - he
  - hi
  - zh
  - yue
  - ja
  - ko
  - th
task_categories:
  - audio-to-audio
  - automatic-speech-recognition
tags:
  - neural-codec
  - speech-tokenizer
  - benchmark
  - multilingual
  - fleurs
pretty_name: codecbench multilingual v1
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files: manifest.parquet

codecbench-multilingual-v1

Evaluation subset used by the Multilingual Neural Audio Codec Leaderboard: 1200 utterances (3.35 h) from the FLEURS test split in 24 languages, 16 kHz mono FLAC.

Selection is deterministic and codec-independent: 3–15 s, one utterance per sentence id, alternating gender, ranked by sha256("codecbench-v1:{lang}:{id}:{path}"); clipped/silent recordings and recordings whose original Whisper-large-v3 transcript has >50% WER/CER are removed.

Note: FLEURS test splits of some languages contain a single gender (male only: de, pt, nl, cs, tr, id, vi, uk, fa, he; female only: ar, th) and FLEURS has no speaker ids, so speaker diversity per language is limited to the FLEURS test speakers.

Columns: uid, lang, fleurs_config, fleurs_id, fleurs_path, gender, duration, text (raw transcription), text_norm, audio (relative FLAC path), asr_orig (Whisper-large-v3 on the original), err_orig.

Derived from google/fleurs (CC-BY-4.0); please cite FLEURS:

@article{fleurs2022, title={FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech},
  author={Conneau, Alexis and others}, journal={arXiv preprint arXiv:2205.12446}, year={2022}}

Per-language summary

lang language n minutes
ar Arabic 50.0 7.5
cs Czech 50.0 8.9
de German 50.0 8.2
el Greek 50.0 7.5
en English 50.0 8.0
es Spanish 50.0 8.9
fa Persian 50.0 10.0
fr French 50.0 7.8
he Hebrew 50.0 6.9
hi Hindi 50.0 8.3
id Indonesian 50.0 8.6
it Italian 50.0 9.6
ja Japanese 50.0 9.5
ko Korean 50.0 9.4
nl Dutch 50.0 7.2
pl Polish 50.0 7.6
pt Portuguese 50.0 8.8
ru Russian 50.0 8.5
th Thai 50.0 7.6
tr Turkish 50.0 9.1
uk Ukrainian 50.0 8.4
vi Vietnamese 50.0 8.3
yue Cantonese 50.0 8.3
zh Mandarin 50.0 7.9