| --- |
| license: cc0-1.0 |
| language: |
| - cy |
| task_categories: |
| - automatic-speech-recognition |
| pretty_name: Welsh ASR (Common Voice 22, filtered + rebalanced) |
| size_categories: |
| - 10K<n<100K |
| tags: |
| - cy |
| - cy |
| - common-voice |
| - speech |
| --- |
| |
| # Welsh ASR (Common Voice 22, filtered + rebalanced) |
|
|
| Welsh (`cy`) speech for ASR, built from **Mozilla Common Voice 22.0** (CC0) via |
| the open [`fsicoli/common_voice_22_0`](https://huggingface.co/datasets/fsicoli/common_voice_22_0) mirror. Built to fine-tune |
| streaming ASR models (e.g. `nvidia/nemotron-3.5-asr-streaming-0.6b`). |
|
|
| ## Splits |
|
|
| | Split | Clips | Hours | |
| |-------|------:|------:| |
| | train | ? | 50.1 | |
| | dev | ? | 0.8 | |
| | test | ? | 2.0 | |
|
|
| `train` = the official validated train split + the filtered `other` bucket + the **excess |
| `dev`/`test` speakers**: Common Voice's official dev/test are balanced for benchmarking and far |
| too large to hold out when fine-tuning a low-resource language, so they are capped (dev |
| ~0.75h, test ~2.0h) by **whole speakers** and the remainder |
| moved into train. Held-out `dev`/`test` stay **disjoint from train by both speaker and sentence**. |
|
|
|
|
|
|
| ## Columns |
|
|
| `audio` (16 kHz mono), `text`, `lang` (`cy`), `client_id` (anonymized speaker), |
| `split`, `source` (`validated` | `other` | `moved-validated`), `quality_score`. |
|
|
