| --- |
| pretty_name: SSSL annotations |
| --- |
| # SSSL annotations (Lhotse shar cuts, no audio) |
|
|
| These are the Lhotse shar `cuts.NNNNNN.jsonl.gz` manifests for the labelled SSSL training and evaluation |
| data on the Fir cluster. Audio is not included: the `recording.*.tar` shards were left out. Phone |
| alignments are in `supervisions[].alignment.phone` as `[symbol (IPA), start, duration, null]`. |
|
|
| ## Layout |
|
|
| The layout is `{dataset}/{language}.tar`, with one tar for each language of a dataset. Inside a tar, the |
| files keep their source paths: `{dataset}_shar/{split}/cuts.*.jsonl.gz` for the `sssl_data` and |
| `speech_test_shar` sets, and `{dataset}/cuts.*.jsonl.gz` for the training corpora. Common Voice keeps its |
| shard dirs, `commonvoice/{locale}_{locale}_train_N_16k_flac/`. |
|
|
| ```bash |
| hf download charsiu/sssl-annotations librispeech/en.tar --repo-type dataset --local-dir . |
| ``` |
|
|
| ## Statistics |
|
|
| "Hours" is the summed cut duration. "Cuts with phone alignment" counts the cuts that have a non-empty |
| `alignment.phone` tier. The rest are text only, and FLEURS has no alignments at all. |
|
|
| | dataset | languages | tars | cuts | hours | cuts with phone alignment | source root on Fir | |
| |---|---|--:|--:|--:|--:|---| |
| | `wenetspeech` | `zh` | 1 | 14,509,578 | 9,870.0 | 14,509,578 (100.0%) | `speech_alignment_neo_shar` | |
| | `commonvoice` | 59 (see folder) | 59 | 5,514,602 | 8,144.2 | 5,514,602 (100.0%) | `speech_alignment_shar` | |
| | `peoples_speech` | `en` | 1 | 1,496,210 | 5,874.9 | 1,496,210 (100.0%) | `speech_alignment_neo_shar` | |
| | `libriheavy_medium` | `en` | 1 | 1,101,040 | 4,533.3 | 1,099,957 (99.9%) | `speech_alignment_neo_shar` | |
| | `mls_german_train` | `de` | 1 | 469,942 | 1,966.5 | 469,443 (99.9%) | `speech_alignment_neo_shar` | |
| | `dutch_mls` | `nl` | 1 | 374,287 | 1,554.2 | 190,301 (50.8%) | `speech_alignment_neo_shar` | |
| | `mls_french_train` | `fr` | 1 | 258,213 | 1,076.6 | 257,910 (99.9%) | `speech_alignment_neo_shar` | |
| | `librispeech` | `en` | 1 | 286,514 | 971.7 | 286,514 (100.0%) | `sssl_data` | |
| | `mls_spanish_train` | `es` | 1 | 220,701 | 917.7 | 217,899 (98.7%) | `speech_alignment_neo_shar` | |
| | `fleurs` | 77 (see folder) | 77 | 159,853 | 544.0 | 0 (0.0%) | `sssl_data` | |
| | `italian_mls_ipa` | `it` | 1 | 59,623 | 247.4 | 58,286 (97.8%) | `speech_alignment_neo_shar` | |
| | `mls_portuguese_train` | `pt` | 1 | 37,533 | 161.0 | 37,123 (98.9%) | `speech_alignment_neo_shar` | |
| | `commonphone` | `de`, `en`, `es`, `fr`, `it`, `ru` | 6 | 76,307 | 116.5 | 76,307 (100.0%) | `sssl_data` | |
| | `mls_polish_train` | `pl` | 1 | 25,043 | 103.6 | 24,772 (98.9%) | `speech_alignment_neo_shar` | |
| | `russian_librispeech` | `ru` | 1 | 54,472 | 92.8 | 54,445 (100.0%) | `speech_alignment_neo_shar` | |
| | `aishell3` | `zh` | 1 | 88,035 | 85.6 | 88,035 (100.0%) | `speech_alignment_neo_shar` | |
| | `zeroth_korean` | `ko` | 1 | 22,720 | 52.9 | 22,640 (99.6%) | `speech_alignment_neo_shar` | |
| | `wiktionary` | `ces`, `deu`, `eng`, `fra`, `pol`, `rus`, `swe`, `ukr` | 8 | 99,596 | 40.5 | 99,596 (100.0%) | `sssl_data` | |
| | `thchs30` | `zh` | 1 | 13,388 | 34.2 | 13,388 (100.0%) | `sssl_data` | |
| | `uk_ireland_english` | `en` | 1 | 17,877 | 31.3 | 17,869 (100.0%) | `speech_alignment_neo_shar` | |
| | `google_spanish` | `es` | 1 | 18,698 | 29.8 | 18,698 (100.0%) | `speech_alignment_neo_shar` | |
| | `multilingual_keywords` | `cs`, `de`, `en`, `es`, `ja`, `pt`, `tr` | 7 | 249,433 | 26.9 | 249,422 (100.0%) | `speech_alignment_neo_shar` | |
| | `gtimit_treebank` | `en` | 1 | 11,993 | 24.1 | 11,993 (100.0%) | `sssl_data` | |
| | `doreco` | 38 (see folder) | 38 | 14,369 | 14.8 | 14,369 (100.0%) | `speech_test_shar` | |
| | `buckeye` | `en` | 1 | 19,086 | 12.6 | 19,086 (100.0%) | `sssl_data` | |
| | `african_french` | `fr` | 1 | 10,239 | 12.5 | 10,026 (97.9%) | `speech_alignment_neo_shar` | |
| | `gtimit_simple` | `en` | 1 | 11,983 | 12.3 | 11,983 (100.0%) | `sssl_data` | |
| | `gtimit_thai` | `th` | 1 | 6,000 | 11.8 | 6,000 (100.0%) | `sssl_data` | |
| | `media_speech_turkish` | `tr` | 1 | 2,513 | 10.0 | 2,506 (99.7%) | `speech_alignment_neo_shar` | |
| | `alffa_swahili` | `sw` | 1 | 10,180 | 9.9 | 1,246 (12.2%) | `speech_alignment_neo_shar` | |
| | `nigerian_english` | `en` | 1 | 3,359 | 5.8 | 3,359 (100.0%) | `speech_alignment_neo_shar` | |
| | `timit` | `en` | 1 | 6,300 | 5.4 | 6,300 (100.0%) | `sssl_data` | |
| | `voxangeles` | 95 (see folder) | 95 | 5,445 | 1.6 | 5,445 (100.0%) | `sssl_data` | |
| | **total** | | 316 | 25,255,132 | 36,596.4 | 24,895,308 (98.6%) | | |
|
|
|
|
| ## Notes |
|
|
| - **Language codes** are each dataset's own: ISO 639-1 (`en`, `de`), Common Voice locales (`zh-CN`, `fy-NL`), |
| FLEURS locales (`th_th`), ISO 639-3 (Wiktionary, VoxAngeles) and Glottocodes (DoReCo). LibriSpeech, TIMIT, |
| Buckeye, gTIMIT simple/treebank and the English training corpora are under `en`; THCHS-30, AISHELL-3 and |
| WenetSpeech are under `zh`. |
| - **Byte-identical copies**, except for VoxAngeles. Its 3 original shards mix 95 languages, so they were split |
| into one `cuts.000000.jsonl.gz` per language, keeping the cut order. |
| - **Duplicate copies left out:** the `speech_alignment_neo_shar` copies of LibriSpeech, Common Phone and |
| THCHS-30, and the `speech_test_shar` copies of the evaluation sets. They are byte-identical to the files here. |
| - **Not included:** the unlabelled pretraining corpora (`speech_unsup_shar`). |
|
|
|
|
