Speech Salamandra ES–EN
Speech Salamandra ES–EN combines spoken input with text instructions for transcription and spoken question answering.
Model and training
The model joins Salamandra 7B Instruct with Meta's SeamlessM4T v2 speech encoder. Its inference flow can produce an intermediate transcript and reasoning before the final text answer. The provided skip_thinking option hides intermediate output; it is not a separately trained model. The release does not establish speech synthesis.Speech Salamandra ES–EN: official model card
Practical use
Practical guidance. Evaluate the transcript and final answer as separate outputs. This helps distinguish a recognition failure from a reasoning error. Include spoken questions whose answers are absent from the recording, so that fluent completion does not look like successful listening. An audio-input interface alone does not make a model a voice generator.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | speech-salamandra-es-en |
| License | Apache License 2.0.Speech Salamandra ES–EN: official model card |
| Model size | 7B language backbone with a SeamlessM4T v2 speech encoder.Speech Salamandra ES–EN: official model card |
Related models
The original Salamandra 7B Instruct is BSC's first conversational adaptation of its 7B multilingual base.
Launch date not confirmedResearch releaseDownloadable- Salamandra TAV 7B
Barcelona Supercomputing Center · Salamandra TAV
Salamandra TAV 7B adds speech input to BSC's translation model for transcription and translation into text.
Launch date not confirmedDownloadable