Datasets:
banc-trawsgrifiadau-bangor-translations
English translations of the verbatim Welsh sentence column of
techiaith/banc-trawsgrifiadau-bangor, produced with the
techiaith/bergamot-mt-cy-en
tiny on-device cy→en student model.
Each Welsh sentence is translated twice:
translated-verbatim— the raw verbatim speech transcription is translated as-is.translated-normalized— the Welsh is first normalized into standard written Welsh with the techiaith-speech-normalizer, then translated.
The goal is to measure what difference Welsh speech normalization makes to the downstream English translation.
Columns
| Column | Description |
|---|---|
split |
Source split (train / validation / test) |
sentence |
Verbatim Welsh transcription (source) |
normalized |
Welsh after speech normalization |
translated-verbatim |
English from the verbatim Welsh |
translated-normalized |
English from the normalized Welsh |
Splits
| Split | Rows |
|---|---|
| train | 48,611 |
| validation | 3,901 |
| test | 3,899 |
Headline numbers
- 56,411 sentences translated.
- Normalization changed the Welsh in 43,906 rows (77.8%).
- The English translation differs between the two routes in 41,382 rows (73.4%).
See analysis_report.md for the full breakdown and banc-trawsgrifiadau-bangor-translations.tsv for a flat TSV
of all results. Method: see EXPERIMENT_METHOD.md in the
source project.
- Downloads last month
- 6