Salamandra TAV 7B
Salamandra TAV 7B adds speech input to BSC's translation model for transcription and translation into text.
Model and training
BSC combines SalamandraTA 7B Instruct with the slprl/mhubert-base-25hz speech encoder. Training includes 12,000 hours of recognition data, 900 hours of speech translation and 113 million text-translation target tokens. The card lists Asturian, Basque, Catalan, English, Galician, Portuguese and Spanish. Outputs are text; this is not a speech synthesiser.Salamandra TAV 7B: official model card
Practical use
Practical guidance. Separate recognition mistakes from translation mistakes when assessing this model. Compare a direct speech translation with translation of a corrected transcript to locate the error source. Include accented speech, short utterances and names. Use a synthesis model only if the workflow also needs spoken output.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | salamandra-TAV-7b |
| License | Apache License 2.0.Salamandra TAV 7B: official model card |
| Model size | 7B translation backbone plus speech encoder.Salamandra TAV 7B: official model card |
Related models
- SalamandraTA 7B v1
Barcelona Supercomputing Center · SalamandraTA
SalamandraTA 7B v1 is the first versioned 7B translation release based on BSC's multilingual Salamandra model.
Launch date not confirmedDownloadable - Speech Salamandra ES–EN
Barcelona Supercomputing Center · Speech Salamandra
Speech Salamandra ES–EN combines spoken input with text instructions for transcription and spoken question answering.
Launch date not confirmedDownloadable