Sarvam AI has released Saaras V4, its latest speech recognition model, with support for 22 Indian languages and English. The model is designed to process speech in different formats, including verbatim transcription, normalised text, code-mixed text, transliteration and translation.
Five speech output modes
Saaras V4 can generate five forms of output from the same audio. Sarvam AI said the model handles these formats within the model rather than relying on separate post-processing systems.
The company said, “Saaras V4 handles these requirements within the model itself. From the same audio input, it can produce five transcript modes: verbatim, normalised, code-mixed, transliterated, and translated.”
The verbatim mode retains the speech as spoken. Transcribe produces text in the native script while normalising numbers and dates. Codemix retains English words in English. Translit converts the speech into English script, while Translate produces an English translation.
Sarvam AI said the five formats can be generated from one model, avoiding additional processing stages. “Because all five come from one model, there are no extra pre-processing steps that could lead to cascading of errors,” the company said.
Built for Indian-language speech
Saaras V4 combines an audio encoder with a 3-billion-parameter hybrid state-space language model that Sarvam AI trained in-house from scratch. The company said the system is designed to handle code-mixing, dialect variations and noisy audio.
Sarvam AI said Saaras V4 recorded a 5.22 per cent language-identification error rate across all 22 Indian languages using verified IndicVoices data. The error rate was 2.9 per cent across the 10 most widely spoken Indian languages.
The company also said Saaras V4 achieved “SOTA performance on all 22 Indian languages” in its evaluations.
Performance across English benchmarks
Sarvam AI evaluated Saaras V4 across seven English speech-recognition benchmarks covering Indian English, international accents, meetings, media, finance and other real-world speech settings.
The company said, “Across these datasets, Saaras V4 achieves the lowest average word error rate.” The datasets include AMI for meeting-room speech, GigaSpeech for podcasts, audiobooks and web video, LibriSpeech, SPGISpeech for financial calls, VoxPopuli and Svarah for Indian-accented English.
For Indian languages, Sarvam AI evaluated the model on the Vistaar benchmark across 10 languages. It used both Word Error Rate (WER) and LLM-WER. Sarvam AI describes WER as the standard measure for speech recognition, while LLM-WER adds a semantic assessment to determine whether a transcription difference changes the meaning of the spoken content.
Streaming and long-form audio
Saaras V4 supports real-time speech streaming. Sarvam AI said the model has a “time to first token below 150 ms” for streaming applications.
The company also said the model can process multi-minute recordings in under a second.
Saaras V4 is available through Sarvam AI’s API. Developers can use Python and Node.js SDKs, while the company has integrations for Vercel AI SDK, LiveKit Agents and Pipecat Agents.
Leave a Reply