Sarvam's Saaras v3 Powers Equal AI's 285M Monthly Indian Calls

Answering the phone in India is a linguistic minefield. A single caller might slip between Hindi, English, and Tamil inside one sentence, on an 8 kHz telephony line, with a background of traffic or a call center behind them. That is the environment where Equal AI's Personal Call Assistant operates, and it now runs its entire speech-to-text stack on Sarvam.

The assistant picks up incoming calls on behalf of users, figures out who is calling and why, then hands back a summary and a recording once the call ends. It handles over a million live calls every day across nine Indian languages, according to Equal CEO Akhilesh Damaraju. Under the hood, every one of those calls is transcribed by Saaras v3, Sarvam's flagship speech recognition model.

Why an India-first ASR was the only option

Global speech models tend to buckle on Indian audio. Hindi speakers routinely switch to English and back within a single sentence, Tamil is agglutinative so word boundary detection is much harder, and Indian telephony audio is often 8kHz with heavy background noise, nothing like the clean studio recordings global ASR models are optimized for. A call assistant that misfires on any of these fails silently, mislabeling who called and why.

Saaras v3 was built specifically for this mess. It was trained on over 1 million hours of real Indian speech through a 4-stage pipeline of large-scale pre-training, supervised fine-tuning, reinforcement learning, and post-training for long-tail errors, and hits 19.31% WER on the IndicVoices benchmark across 10 languages, down from 22% in v2. The model covers 23 languages, 22 Indian plus English.

Inside the production setup

Equal AI is not just using vanilla transcription. The workload spans multiple output modes including transcription, translation, verbatim, transliteration, and code-mixed outputs, plus real-time streaming, batch, and speaker diarization for multi-party calls.

The most interesting choice is which mode dominates. Transliteration, converting spoken Indian-language audio into Roman script, accounts for roughly 97% of Equal AI's speech recognition calls. The reason is downstream compatibility: Roman-script text drops cleanly into existing financial workflow systems that were never designed to parse Devanagari or Tamil script. Here is what the call looks like in code:

from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_KEY")
response = client.speech_to_text.transcribe(
    file_path="call.wav",
    language="hi-IN",
    model="saaras:v3"
)
print(response.transcript)

The numbers that matter

The case study lays out a growth curve that is unusual even by API-startup standards:

• Volume: Approximately 34 million minutes of audio processed every month on Sarvam.
• Growth: Production volume grew roughly 420x between October 2025 and May 2026, from about 555,000 calls to 234 million calls.
• Run rate: By June 2026, usage reached approximately 285 million calls per month.
• Reliability: A 99.98% success rate across roughly 129 million calls in the most recent 30-day window.
• Model mix: Saaras v3 powers about 85% of Equal AI's speech recognition traffic, with older versions handling the rest.
• Monthly cadence: Triple-digit month-on-month growth from November through March, moderating to 38% in April and 14% in May as the platform hit a steady high-volume plateau.

What this signals for Indic voice AI

Two things stand out. First, Indian-language ASR is finally clearing the enterprise bar. A 99.98% success rate at nine-figure monthly call volumes is not a demo, it is a production system that fintech customers are betting workflows on. Second, the transliteration-heavy usage pattern is a tell about how Indic AI actually gets deployed in the wild: the LLMs and business systems consuming the transcripts still expect Latin script, so the ASR layer is doubling as a script normalizer.

For anyone building voice agents targeting India, there are a few practical takeaways. Saaras v3 quotes sub-150ms time to first token with configurable Accurate, Balanced, and Fast modes, which is workable for live call screening. All audio is processed and stored in India with no cross-border transfers and full compliance with Indian data regulations, which matters for anyone touching regulated data like KYC or lending calls. And the model exposes distinct modes for verbatim capture versus clean transcription versus transliteration, so you can match the output shape to whatever your downstream pipeline actually needs rather than post-processing the same transcript five different ways.

The Equal deployment is essentially the first public proof point that a domestically built Indic ASR can run a nine-language, million-call-a-day consumer product without collapsing on latency or accuracy. That is a different conversation than the one Indian voice AI was having a year ago.