---
license: other
license_name: cc-by-3.0-and-cc-by-4.0
license_link: LICENSE
pretty_name: PolyPlex Word Timings
tags:
- forced-alignment
- word-timestamps
- speech
- multilingual
configs:
- config_name: fleurs
  data_files:
  - split: train
    path: "fleurs/*.parquet"
- config_name: yodas2
  data_files:
  - split: train
    path: "yodas2/*.parquet"
---

# PolyPlex Word Timings

Word-level onset and end times for multilingual speech, obtained by forced alignment with Whisper large-v3. Each row is one
aligned piece of speech: its audio (16 kHz mono FLAC, playable in the dataset viewer), its transcript, and the onset and end
of every word.

The timings were made to measure, for the PolyPlex full-duplex speech model, how often a word's text tokens do not fit the
80 ms frames before the next word starts. They are useful wherever word timings for real speech in many languages are needed.

## Configurations

| Config | Source | Speech | License |
|---|---|---|---|
| `yodas2` | [YODAS2](https://huggingface.co/datasets/espnet/yodas2), user-uploaded subtitle subsets (`<lang>000`) | Real speech from YouTube videos under Creative Commons | CC BY 3.0 |
| `fleurs` | [FLEURS](https://huggingface.co/datasets/google/fleurs), first 200 utterances of `dev` | Read sentences | CC BY 4.0 |

### `yodas2`: real speech, 84 languages

| Language | Source | Pieces | Words |
|---|---|---|---|
| Afrikaans | `af000` | 72 | 2,071 |
| Armenian | `hy000` | 421 | 7,031 |
| Assamese | `as000` | 104 | 946 |
| Bashkir | `ba000` | 14 | 73 |
| Basque | `eu000` | 1,000 | 33,176 |
| Belarusian | `be000` | 1,000 | 44,047 |
| Bengali | `bn000` | 472 | 8,901 |
| Bhojpuri | `bh000` | 1 | 48 |
| Bosnian | `bs000` | 801 | 28,728 |
| Bulgarian | `bg000` | 753 | 22,385 |
| Burmese | `my000` | 299 | 2,747 |
| Catalan | `ca000` | 1,000 | 38,587 |
| Croatian | `hr000` | 1,000 | 30,213 |
| Czech | `cs000` | 1,000 | 30,688 |
| Danish | `da000` | 1,000 | 40,163 |
| Dutch | `nl000` | 1,000 | 41,378 |
| Eastern Yiddish | `yi000` | 8 | 275 |
| English | `en000` | 1,000 | 43,379 |
| Estonian | `et000` | 1,000 | 31,981 |
| Faroese | `fo000` | 10 | 217 |
| Finnish | `fi000` | 1,000 | 29,147 |
| French | `fr000` | 1,000 | 46,976 |
| Galician | `gl000` | 519 | 17,362 |
| Georgian | `ka000` | 1,000 | 36,375 |
| German | `de000` | 785 | 25,801 |
| Greek | `el000` | 1,000 | 33,237 |
| Gujarati | `gu000` | 465 | 5,065 |
| Haitian | `ht000` | 143 | 3,655 |
| Hebrew | `iw000` | 1,000 | 41,621 |
| Hindi | `hi000` | 1,000 | 36,099 |
| Hungarian | `hu000` | 1,000 | 17,634 |
| Icelandic | `is000` | 654 | 24,036 |
| Indonesian | `id000` | 1,000 | 25,885 |
| Irish | `ga000` | 182 | 5,181 |
| Italian | `it000` | 1,000 | 38,902 |
| Japanese | `ja000` | 770 | 4,617 |
| Javanese | `jv000` | 202 | 3,217 |
| Kannada | `kn000` | 591 | 9,093 |
| Kazakh | `kk000` | 458 | 7,659 |
| Khmer | `km000` | 444 | 3,276 |
| Korean | `ko000` | 1,000 | 21,657 |
| Lao | `lo000` | 25 | 126 |
| Latvian | `lv000` | 328 | 10,414 |
| Lithuanian | `lt000` | 1,000 | 27,831 |
| Luxembourgish | `lb000` | 9 | 246 |
| Macedonian | `mk000` | 787 | 25,747 |
| Malay | `ms000` | 1,000 | 15,555 |
| Malayalam | `ml000` | 975 | 24,989 |
| Marathi | `mr000` | 575 | 9,364 |
| Nepali | `ne000` | 556 | 7,451 |
| North Azerbaijani | `az000` | 799 | 17,138 |
| Northern Uzbek | `uz000` | 830 | 19,177 |
| Norwegian Bokmal | `no000` | 838 | 35,707 |
| Occitan | `oc000` | 28 | 1,000 |
| Oriya | `or000` | 297 | 1,870 |
| Persian | `fa000` | 797 | 29,764 |
| Polish | `pl000` | 994 | 30,952 |
| Portuguese | `pt000` | 972 | 44,293 |
| Punjabi | `pa000` | 467 | 7,612 |
| Romanian | `ro000` | 1,000 | 32,751 |
| Russian | `ru000` | 1,000 | 30,844 |
| Sardinian | `sc000` | 2 | 40 |
| Serbian | `sr000` | 1,000 | 32,721 |
| Simplified Chinese | `zh000` | 860 | 7,567 |
| Sindhi | `sd000` | 13 | 155 |
| Sinhala | `si000` | 923 | 13,833 |
| Slovak | `sk000` | 1,000 | 26,622 |
| Slovenian | `sl000` | 1,000 | 24,210 |
| Spanish | `es000` | 766 | 27,778 |
| Standard Arabic | `ar000` | 1,000 | 23,562 |
| Sundanese | `su000` | 13 | 113 |
| Swahili | `sw000` | 633 | 14,926 |
| Swedish | `sv000` | 1,000 | 33,230 |
| Tajik | `tg000` | 47 | 1,128 |
| Tamil | `ta000` | 974 | 27,073 |
| Tatar | `tt000` | 2 | 3 |
| Telugu | `te000` | 698 | 9,096 |
| Thai | `th000` | 883 | 8,976 |
| Tosk Albanian | `sq000` | 573 | 20,474 |
| Turkish | `tr000` | 1,000 | 23,371 |
| Ukrainian | `uk000` | 913 | 32,020 |
| Urdu | `ur000` | 638 | 21,633 |
| Vietnamese | `vi000` | 837 | 32,817 |
| Welsh | `cy000` | 487 | 23,310 |

### `fleurs`: read speech, 79 languages

| Language | Source | Pieces | Words |
|---|---|---|---|
| Afrikaans | `af_za` | 198 | 4,197 |
| Armenian | `hy_am` | 199 | 3,802 |
| Assamese | `as_in` | 192 | 3,283 |
| Asturian | `ast_es` | 200 | 4,143 |
| Belarusian | `be_by` | 199 | 4,039 |
| Bengali | `bn_in` | 191 | 3,291 |
| Bosnian | `bs_ba` | 199 | 3,803 |
| Bulgarian | `bg_bg` | 199 | 4,103 |
| Burmese | `my_mm` | 116 | 906 |
| Cantonese | `yue_hant_hk` | 199 | 313 |
| Catalan | `ca_es` | 199 | 4,624 |
| Cebuano | `ceb_ph` | 196 | 4,605 |
| Croatian | `hr_hr` | 200 | 3,487 |
| Czech | `cs_cz` | 200 | 3,459 |
| Danish | `da_dk` | 199 | 3,955 |
| Dutch | `nl_nl` | 171 | 3,695 |
| English | `en_us` | 200 | 3,884 |
| Estonian | `et_ee` | 200 | 3,050 |
| Finnish | `fi_fi` | 200 | 3,135 |
| French | `fr_fr` | 200 | 4,748 |
| Galician | `gl_es` | 200 | 4,468 |
| Georgian | `ka_ge` | 175 | 2,459 |
| German | `de_de` | 198 | 3,793 |
| Greek | `el_gr` | 200 | 4,474 |
| Gujarati | `gu_in` | 188 | 3,622 |
| Hebrew | `he_il` | 200 | 3,370 |
| Hindi | `hi_in` | 200 | 4,530 |
| Hungarian | `hu_hu` | 198 | 3,558 |
| Icelandic | `is_is` | 36 | 682 |
| Indonesian | `id_id` | 200 | 3,684 |
| Irish | `ga_ie` | 200 | 4,641 |
| Italian | `it_it` | 197 | 4,257 |
| Japanese | `ja_jp` | 200 | 203 |
| Javanese | `jv_id` | 199 | 3,807 |
| Kabuverdianu | `kea_cv` | 196 | 4,180 |
| Kannada | `kn_in` | 199 | 2,839 |
| Kazakh | `kk_kz` | 197 | 2,968 |
| Khmer | `km_kh` | 188 | 871 |
| Korean | `ko_kr` | 200 | 3,040 |
| Lao | `lo_la` | 161 | 674 |
| Latvian | `lv_lv` | 200 | 3,314 |
| Lithuanian | `lt_lt` | 200 | 3,249 |
| Luxembourgish | `lb_lu` | 200 | 3,994 |
| Macedonian | `mk_mk` | 200 | 3,829 |
| Malay | `ms_my` | 200 | 3,764 |
| Malayalam | `ml_in` | 181 | 2,288 |
| Maltese | `mt_mt` | 199 | 3,483 |
| Marathi | `mr_in` | 199 | 3,598 |
| Nepali | `ne_np` | 199 | 3,329 |
| North Azerbaijani | `az_az` | 198 | 3,562 |
| Northern Uzbek | `uz_uz` | 200 | 3,262 |
| Norwegian Bokmal | `nb_no` | 161 | 3,350 |
| Occitan | `oc_fr` | 199 | 4,352 |
| Oriya | `or_in` | 167 | 2,745 |
| Persian | `fa_ir` | 196 | 4,179 |
| Polish | `pl_pl` | 200 | 3,456 |
| Portuguese | `pt_br` | 200 | 4,323 |
| Punjabi | `pa_in` | 199 | 4,392 |
| Romanian | `ro_ro` | 200 | 4,514 |
| Russian | `ru_ru` | 200 | 3,601 |
| Serbian | `sr_rs` | 200 | 3,672 |
| Simplified Chinese | `cmn_hans_cn` | 200 | 403 |
| Sindhi | `sd_in` | 200 | 4,778 |
| Slovak | `sk_sk` | 200 | 3,651 |
| Slovenian | `sl_si` | 200 | 3,667 |
| Spanish | `es_419` | 200 | 4,735 |
| Standard Arabic | `ar_eg` | 199 | 3,617 |
| Swahili | `sw_ke` | 199 | 3,914 |
| Swedish | `sv_se` | 200 | 3,956 |
| Tagalog | `fil_ph` | 192 | 4,399 |
| Tajik | `tg_tj` | 195 | 4,132 |
| Tamil | `ta_in` | 197 | 2,869 |
| Telugu | `te_in` | 194 | 2,987 |
| Thai | `th_th` | 198 | 745 |
| Turkish | `tr_tr` | 200 | 3,133 |
| Ukrainian | `uk_ua` | 199 | 3,615 |
| Urdu | `ur_pk` | 200 | 5,090 |
| Vietnamese | `vi_vn` | 199 | 5,612 |
| Welsh | `cy_gb` | 198 | 4,338 |

## Columns

| Column | Meaning |
|---|---|
| `audio` | The aligned audio, 16 kHz mono FLAC; word times count from its first sample |
| `language` | Language name |
| `subset` / `fleurs_config` | Source subset (YODAS2) or FLEURS configuration |
| `video_id`, `piece_start` | (`yodas2`) The source video and the start of the piece in it, in seconds |
| `id` | (`fleurs`) FLEURS utterance file name without extension |
| `text` | Transcript the alignment was forced to |
| `duration` | Length of `audio`, seconds |
| `words` | `text` split on whitespace |
| `word_starts`, `word_ends` | Onset and end of each word, seconds from the start of the piece |
| `alignment_loss` | Mean teacher-forced cross entropy of the transcript tokens under Whisper; high values mean the transcript does not match the speech |

## How it was made

- **Alignment.** The reference transcript is fed to the Whisper large-v3 decoder under teacher forcing. The cross attention
  of the checkpoint's alignment heads is normalized per head and median filtered, and dynamic time warping over the token by
  frame matrix gives each token the frame it starts on (20 ms resolution). Recognition errors do not move the words, because
  the transcript is not decoded.
- **YODAS2 pieces.** For each language the subtitle file of four shards spread over the subset is read, and the audio shard is
  streamed until 250 pieces per shard are aligned. Neighbouring subtitle segments of one video are joined into pieces of at
  most 25 seconds, split at pauses over one second, with 0.3 seconds of margin on both sides.
- **Left out.** Clips longer than 30 seconds (Whisper's window) and transcripts longer than 448 decoder tokens.
- **Languages Whisper lacks** are prompted with a close language (Asturian as Spanish, Irish as English, Oriya as Bengali,
  Cebuano as Tagalog, Kabuverdianu as Portuguese, Bhojpuri as Hindi, Sardinian as Italian).

## Recommended filtering

YODAS2 subtitles are uploaded by users and are not always a verbatim transcript; a piece whose subtitle does not match the
speech gets compressed, wrong word times. Keep pieces with `alignment_loss <= 1.61`, the 95th percentile of the
loss on human-transcribed Korean conversation (KsponSpeech evaluation split). On Korean, this kept 82% of YODAS2 pieces and
brought the share of words cut at 12.5 Hz to 1.12%, against 0.89% on KsponSpeech and 8.44% before filtering.

## Limitations

- Whisper attention alignment is not human-verified. It sometimes gives a short word almost no duration (onsets of two
  words 0 to 60 ms apart).
- YODAS2 content is whatever the subset holds: conversation, vlogs, lectures, news reading.
- Scripts written without spaces (Chinese, Japanese, Thai, Lao, Khmer, Burmese) are aligned per whitespace unit, which is
  often a whole phrase.

## Licenses and attribution

- `yodas2` is derived from YODAS2 (CC BY 3.0) and carries pieces of its audio. The audio belongs to the creators of the
  YouTube videos, which YODAS collected under Creative Commons licenses. Cite: Li et al., *YODAS: YouTube-Oriented Dataset for Audio and Speech*,
  ASRU 2023.
- `fleurs` is derived from FLEURS (CC BY 4.0) and carries its audio. Cite: Conneau et al., *FLEURS: Few-shot Learning Evaluation of Universal
  Representations of Speech*, SLT 2022.
- The word times added here are released under the license of their source configuration.
- Alignment model: Whisper large-v3 (MIT).

If you appear in a source video and want its rows removed, open a discussion on this repository; rows removed from YODAS2
on request are removed here too.