Conformer Transducer Languages of Spain
BSC's LoS Conformer adapts NVIDIA's Spanish Conformer Transducer to four languages of Spain.
Model and training
The parent is nvidia/stt_es_conformer_transducer_large. BSC fine-tuned it on about 2,700 hours of Catalan, Spanish, Galician and Basque speech. The model has roughly 120 million parameters and produces lowercase, unpunctuated text. The card uses NVIDIA NeMo for inference. BSC's contribution is multilingual training, not the original Conformer architecture.Conformer Transducer Languages of Spain: official model card
Practical use
Practical guidance. This smaller recogniser is relevant when serving resources constrain a Whisper deployment. Compare word errors for each language and record the audio segmentation used. Keep the recogniser's plain output separate from any punctuation or text-normalisation stage added by the application.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | stt_los_conformer_transducer_large |
| License | Apache License 2.0.Conformer Transducer Languages of Spain: official model card |
| Model size | Approximately 120 million parameters.Conformer Transducer Languages of Spain: official model card |
Related models
- Conformer Transducer Languages of Spain Punctuated
Barcelona Supercomputing Center · BSC LoS Conformer
This BSC LoS Conformer release adds trained punctuation to multilingual speech recognition.
Launch date not confirmedDownloadable Whisper Large v3 LoS is BSC's multilingual adaptation for Spanish, Catalan, Galician and Basque speech.
Launch date not confirmedDownloadable