National : SquadStack.ai, which builds and runs outcome-driven Voice AI Sales Agents for India’s leading consumer brands, today launched Arth V2, a streaming speech-to-text model designed for Indian sales calls. Arth is the first model from SquadStack.ai’s model factory. On the SquadStack Conversational Streaming ASR Benchmark (8 kHz), Arth V2 recorded a semantic word error rate of 8.85%, ahead of India’s Sarvam and global players like Deepgram, Google, OpenAI, Cartesia and ElevenLabs, with the lowest latency of any engine tested.
“Speech AI improves so quickly that a model which ships once and stays frozen is out of date within a month,” said Vikas Gulati, Co-founder and CTO, SquadStack.ai. “That is why we run SquadStack.ai as a model factory, and ship a new version whenever it beats what is in production. Lift, our self-improvement loop, feeds Arth every call where a word was misheard, so each version learns from the last one’s mistakes. Expect a new Arth version every few weeks.”
Arth V2 makes 18% fewer meaning-changing errors than Sarvam Saaras v4, 26% fewer than Google Gemini 3.5 Transcribe, 28% fewer than OpenAI gpt-live-transcribe and 60% fewer than ElevenLabs Scribe v2 Realtime. It delivered its final transcript within 67 milliseconds of the customer stopping on 80% of turns, 2.6 times quicker than Sarvam and 12 times quicker than OpenAI. Soniox v5 recorded a slightly lower error rate, at 8.77%, but has a higher latency. No engine tested beat Arth V2 on accuracy and speed together. Arth V2 also got fewer numbers wrong than Sarvam, Deepgram, Google and OpenAI, whether a loan amount, an EMI date or a pincode, which is exactly where Indian customers tend to switch language.
On one benchmark call, a customer quotes “chhah hazaar nau sau hai”, ₹6,900. Google Gemini wrote 1,900, Sarvam wrote “chaar nau sau” and OpenAI wrote just “nau sau”, while Arth V2 got the amount right. One misheard word like that changes what the agent does next.
The benchmark is built from unscripted calls in which customers pause, talk over the agent and switch languages mid-sentence. More than four in five minutes of speech are code-mixed and close to nine in ten carry background noise. The audio is 8 kHz narrowband telephony, which cuts everything above roughly 3.4 kHz, and every engine is tested in streaming mode, listening as the customer speaks. Scoring counts only the mistakes that change what the customer meant. Most Indian-language speech benchmarks rely on clean, wideband recordings. The call set was rebuilt so that none of its calls or speakers overlap with the audio used to train Arth.
Arth was trained in-house, from scratch, on 60 crore+ minutes of Indian sales conversations from more than 10 crore unique speakers, covering over 85% of the country’s pincodes and most of its dialects and accents. It now runs 80% of the 50 lakh+ AI calls SquadStack.ai makes each day.
“Every sales conversation starts with hearing the customer correctly,” said Apurv Agrawal, Co-founder and CEO, SquadStack.ai. “Speech-to-text looks solved, but it is still the biggest point of failure for a voice AI agent. Feed a language model the wrong words and it gives the wrong answer. A customer in Patna and a customer in Ludhiana both speak Hindi, but they sound nothing alike, and both will switch to English halfway through a sentence. We have run AI-native contact centres for India’s leading consumer brands for over six years, and Arth is built on what those calls taught us. Owning our models means every call we run makes the next one better.”
Arth runs inside SquadStack.ai’s Voice AI Agents for all of its enterprise customers. Arth V2 is now also available in closed beta for teams that want to test it on their own calls. The benchmark’s test audio, reference transcripts and scoring method are available on Hugging Face for any team that wants to test its own engine.
****
