Saaras V4
Saaras V4 pairs an audio encoder with a 3B hybrid state-space decoder trained by Sarvam. Its five output modes produce normalized, verbatim, transliterated, code-mixed or English-translated text.
Guide recognition with domain terms
V4 accepts up to 50 keyterms on REST and batch requests. The model guide excludes this parameter from streaming and from V3. Names and technical terms therefore require an endpoint-specific integration.Saaras model guide
Practical use
Practical guidance. Choose transcript form before measuring accuracy. Normalized numbers may suit a searchable record, while verbatim words may suit a dialogue study. Use domain names and difficult recordings to check whether keyterms improve your actual errors.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API |
| Model identifiers | saaras:v4 |
| API position | Current default and recommended Saaras modelSarvam model API catalog |
| Language coverage | 22 Indian languages plus Indian and global EnglishSaaras model guide |
Related models
- Saaras V2.5
Sarvam AI · Saaras
An earlier speech-to-English translation model in the Saaras family.
Launch date not confirmedDeprecated API; check access detailsAPI A speech-recognition model for 22 Indian languages and Indian English with five output modes.
Launch date not confirmedAPI