Shuka 1
Shuka 1 joins Sarvam's Saaras V1 audio encoder to Llama 3 8B Instruct through an approximately 60M-parameter projector. The published model accepts audio and produces text.
Research release
Only the bridge was trained
The encoder and decoder stayed frozen while Sarvam trained the projector on less than 100 hours of English and Hindi audio. The card reports zero-shot checks in other Indian languages.Shuka 1 official model card
Practical use
Practical guidance. Use Shuka to study spoken-question answering as a different task from literal transcription. Score whether it answers the spoken request correctly, and check languages outside the projector's training data separately.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | shuka-1Shuka v1 |
| License | Llama 3Shuka 1 official model card |
Related models
A speech-recognition model for Indian languages and global English with normalized, verbatim, transliterated, code-mixed and translated output.
Launch date not confirmedAPI