| --- |
| license: unknown |
| language: |
| - zh |
| pretty_name: CDC Press Conferences |
| extra_gated_prompt: >- |
| Access requires agreeing to non-commercial research use only. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| Intended use: text |
| I agree to use this dataset only for non-commercial research purposes: checkbox |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-*.parquet |
| dataset_info: |
| features: |
| - name: id |
| dtype: string |
| - name: audio |
| dtype: audio |
| - name: text |
| dtype: string |
| --- |
| |
| # CDC Press Conferences |
|
|
| Utterance-level clips from Taiwan CDC (疾病管制署) press-conference YouTube livestreams |
| (largely COVID-19 era, 2020-2023). **This corpus is Mandarin, not Taiwanese** -- it is one |
| of the few non-Taiwanese-language corpora kept in this project's source bucket, included |
| here for reference/comparison purposes only. |
|
|
| ## Dataset Description |
|
|
| - **Use case**: ASR (Automatic Speech Recognition) |
| - **Language**: Mandarin Chinese (not Taiwanese) |
| - **Source**: Taiwan CDC press-conference YouTube livestreams, 728 episodes across 5 years |
| (2020-2023). Utterance-level audio/text pairs were produced by the team's own prior |
| pipeline (`cdc_press/utils/cdc_preprocess.ipynb` in the source bucket): parse each |
| episode's auto-generated `.srt`, merge consecutive caption entries separated by <100ms |
| gaps (a heuristic fix for captions that get cut mid-word -- see caveat below), then a |
| second merge pass targeting 3.5-22.5 second spans, then crop the corresponding 16kHz WAV |
| per final caption group. This repo reshapes the team's already-built `cdc_ds.zip` (a |
| `datasets.DatasetDict.save_to_disk` export covering all 5 source years) into |
| HF-viewer-friendly Parquet shards; no re-processing of the original video/subtitles was |
| done here. |
| - **Known caveat -- residual mid-word/mid-sentence cuts**: even after the 100ms-gap merge |
| heuristic, some captions may still be cut mid-word or mid-sentence, since the merge is a |
| heuristic, not a real transcript re-alignment. Concrete example from a source `.srt`: |
| one caption ends "...那今天我們指揮", the next begins "中心出席的...", splitting the |
| compound word 指揮中心 ("Central Epidemic Command Center") right through the middle with |
| a 0ms gap between them. Treat `text` as generally reliable but not perfectly clean. |
| - **Not included**: whole-episode raw audio (`16k`/`48k` WAV) and raw `.srt` subtitle files |
| also exist in the source bucket but are not part of this repo -- only the pre-cropped |
| utterance-level version from `cdc_ds.zip` is included here. |
| - **Audio format**: 16kHz, mono, 16-bit PCM WAV |
| - **Split**: this dataset is `train` only (the 5 source years are not modeling splits -- |
| they're just how the raw data happened to be organized during scraping; each row's source |
| episode/date is embedded in `id` -- see Dataset Structure below -- so filter on that if |
| you need a per-year subset). |
|
|
| ## Dataset Structure |
|
|
| - `id`: `{episode_id}_{index:04}`, e.g. `2020-10-14-14_00_7ibcEttw4-g_0001` -- the part |
| before the trailing `_NNNN` utterance index is the source YouTube episode identifier |
| (date + title fragment + video ID). There is no separate episode/origin column since it's |
| always exactly this prefix of `id`. |
| - `audio`: audio clip (16kHz, mono, 16-bit PCM WAV). |
| - `text`: Mandarin transcript for the clip (originally `zh` in the source dataset; same |
| language as the audio, per this project's field-naming convention). |
|
|
| ## Statistics |
|
|
| | **split** | **lang_name** | **hours** | **n_utts** | **n_chars_in_utts** | **secs/utt** | **chars/sec** | **n_sents** | **n_chars_in_sents** | |
| | --- | --- | --- | --- | --- | --- | --- | --- | --- | |
| | train | Mandarin | 404.0792 | 238003 | 7176515 | 6.11 | 4.93 | 0 | 0 | |
| | **Total** | **-** | **404.0792** | **238003** | **7176515** | **6.11** | **4.93** | **0** | **0** | |
| |
