Sarvam AI on Wednesday, unveiled its Saaras V4 automatic speech recognition model, designed to undertake multilingual functions as well as deal with real world audio situations competently parsing multiples sources of sound.
The AI can listen to multiple audio tracks at the same time, tracking each source with notable competence in a single pass. This version also has the ability to ignore background noise and has half the error rate of models such as GPT 4.
The model also has support for Indian languages along with English and make speech transcripts in five different formats. Sarvam achieved these capabilities by integrating an audio encoder with a 3-billion parameter hybrid state-space language model that it also developed.
The company claims the model can take the same audio input and produce verbatim transcripts, normalized text, code-mixed text, transliteration, and translation, all as built-in capabilities, without needing separate post-processing systems. For English, Saaras V4 was tested on seven benchmarks spanning Indian English, international accents, meetings, financial conversations, media and other speech conditions. As per the firm, it achieved the lowest average word error rate across these datasets.
ALSO READ: GPT-6 Sol And Luna Out: OpenAI Launches Cheaper AI Models Amid Calls For 'Slowdown'
Saaras V4 was tested on the Vistaar benchmark for Indian languages, spanning 10 languages via standard Word Error Rate (WER) as well as LLM-WER, whuch adds semantic evaluation to indetntify errors that change meaning and differences caused by differences in formatting or orthography,
It can also automatically identify the spoken language and transcribe it in the matching native script.
Saaras V4 had a WER (Word Error Rate) of 16.03% in the L5 keyword-prompting setting on the IndicContextEval benchmark.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.