Sarvam AI Cuts Transcription Errors 20% for India's Wealth Manager Dezerv

India's wealth management firm Dezerv has published a case study with Sarvam AI detailing how it built a post-call analytics system on Sarvam's Speech to Text API. The numbers make a strong argument for India-first speech models in high-stakes financial contexts.

The insight locked inside every client call

Dezerv's business runs on relationship managers. Every client is assigned one, and RMs explain portfolio decisions, steady clients through volatile markets, and carry the conversations on which long-term trust is built. Those conversations happen over video conferencing, in whichever language the client is most comfortable using.

Dezerv's RM calls were the company's richest source of customer insight and its least accessible. Every call revealed what clients worried about, which explanations landed, and where RMs needed coaching. With thousands of calls every month, all of that was locked inside audio recordings nobody could search.

The previous transcription pipeline compounded the problem. It was outsourced end to end, so Dezerv could not audit its own quality. When the team ran comparisons, the transcripts fell short, especially for Indic languages.

Why generic ASR breaks on Indian speech

The transcription challenge is real. Dezerv's clients and RMs speak the way urban India speaks: mixing English, Hindi, and regional languages within a single sentence, with dense references to fund names and numbers, over video conferencing calls. This code-mixing, switching languages mid-sentence, is something most Western-trained ASR (automatic speech recognition) models handle poorly.

Before committing to Sarvam, Dezerv ran a proper benchmark across multiple leading Speech to Text platforms using open-source datasets, including AI4Bharat's Svarah. The team measured Word Error Rate (WER) and Character Error Rate (CER) across English, Hindi, Marathi, Telugu, and Tamil. WER is the standard transcription accuracy metric: it measures what fraction of words in a transcript are wrong.

Three things separated Sarvam from the field:

• Indian audio training data: Sarvam's models are trained on over a million hours of Indian audio, including code-mixed conversations, regional accents, and telephony recordings.
• Built-in speaker diarization: The model automatically separates speakers throughout a conversation. In wealth management, knowing who said something matters as much as knowing what was said.
• Async batch processing: Dezerv submits large volumes of recordings asynchronously and receives structured, speaker-separated transcripts without building or operating any transcription infrastructure of its own.

What they built

Dezerv integrated Sarvam through its SDK and runs three production use cases: transcription of video meetings, transcription of phone calls, and live note-taking for RMs during client conversations.

Recordings are processed via Sarvam's Batch API with speaker diarization enabled. Here is what a basic transcription call looks like using the Sarvam Speech to Text API:

from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
response = client.speech_to_text.transcribe(
    file_path="client_call.wav",
    language="hi-IN",
    model="saaras:v3"
)
print(response.transcript)
for word in response.words:
    print(f"[{word.start:.2f}s] {word.text}")

On data governance: client data is managed under Dezerv's own framework, and Zero-Retention AI Processing ensures no client-identifying information is stored with Sarvam. For a SEBI-regulated firm handling portfolios starting at ₹50 lakhs, that is a hard requirement.

Results at scale

Dezerv now processes close to 3.6 lakh minutes of speech through Sarvam every month, roughly 3.2 lakh minutes of video meetings and 40,000 minutes of phone calls. That is around 60,000 hours of conversation per month flowing through a single API.

Word Error Rate dropped by around 20% compared to the previous setup. Bringing the pipeline in-house also gave the team direct visibility into quality, something the outsourced arrangement never offered.

For teams evaluating similar pipelines, responsiveness is worth factoring in alongside accuracy. When Dezerv needed video conferencing transcription, a capability not supported at the start of the engagement, Sarvam added it within weeks. When Dezerv later needed support for calls longer than an hour, the required infrastructure changes shipped in under a week.

Where this fits in Indian fintech

Sarvam AI was founded in 2023 in Bengaluru and was selected under India's IndiaAI Mission to build the country's first homegrown AI stack, covering speech, translation, and large language models optimized for how Indian languages are actually used. The Dezerv deployment sits alongside other enterprise and government rollouts, including UIDAI's Aadhaar voice agents serving 4 crore Indians.

As of June 2026, Dezerv manages over ₹17,000 crore across PMS, AIF, and distribution with a team of 650 across five cities. At that scale, what the firm learns from client conversations compounds directly into business outcomes: where advisors need coaching, what clients are worried about, and how better advice gets delivered across thousands of relationships simultaneously.

A 20% WER reduction sounds like a benchmark footnote until you consider the context. In a compliance-sensitive environment where a misheard number can mean a misunderstood instruction, the gap between generic transcription and a model trained on Indian speech is the gap between a system you can act on and one you cannot.