Sarvam AI has unveiled Saaras V4, its latest automatic speech recognition model, which can process multiple audio sources simultaneously while supporting Indian languages and English.
The model is designed for real-world audio, including background noise and conversations involving more than one speaker, and can produce speech in five formats.
Also Read: Sarvam AI Enters India’s Unicorn Club: Secures $234 Million in HCLTech-Led Funding Round
Saaras V4 Can Track Multiple Audio Sources
Saaras V4 can listen to multiple audio tracks in a single pass and track each source separately. Sarvam said the model can also ignore background noise and has half the error rate of models such as GPT-4.
The system combines an audio encoder with a 3-billion-parameter hybrid state-space language model developed by Sarvam. It can take the same audio and generate verbatim transcripts, normalised text, code-mixed text, transliteration and translation without separate post-processing systems.
For English, Saaras V4 was tested across seven benchmarks covering Indian English, international accents, meetings, financial conversations, media and other speech conditions. Sarvam said it recorded the lowest average word error rate across these datasets.
Also Read: India AI Impact Summit: Sarvam AI Debuts Kaze, Marking Its Entry Into Wearable AI
Indian Language Testing And Speech Recognition
Sarvam V4 was also tested on the Vistaar benchmark for Indian languages, covering 10 languages using standard Word Error Rate, or WER, as well as LLM-WER. The latter adds a semantic check to identify errors that alter meaning, while accounting for differences in formatting and writing conventions.
The model can automatically identify the spoken language and transcribe it in the corresponding native script. On the IndicContextEval benchmark, Saaras V4 recorded a 16.03% WER in the L5 keyword-prompting setting.
Sarvam’s own evaluation says Saaras V4 achieved state-of-the-art performance across all 22 Indian languages and also recorded the lowest WER across seven global English datasets. Its language identification error rate was 2.9% across the top 10 Indian languages and 5.22% across all 22.
For the market, the immediate point to watch is how Saaras V4’s combination of multilingual speech recognition, multiple output formats and handling of noisy, mixed-source audio translates into practical voice applications.
Samarjit Kaur is a journalist and communications professional covering technology & emerging digital trends. With a focus on clarity and context, she reports on developments shaping industries and governance. When not reporting, she chooses to plug-in and relax on her playlists and plan her next bucket-list trips!
