India AI Impact Summit 2026

demystifying-voice-stack-what-makes-voice-ai-work-at-scale

8 min read

TRACK:Democratizing AI Resources
SPEAKERS
  • Ananth Nagaraj, Gnani.ai
  • Debdoot Mukherjee, Meesho
  • Deepika Mogilishetty, EkStep Foundation
  • Maitreya Wagh, Bolna
  • Sunil Gupta, Yotta
SESSION OVERVIEW

This session brings together practitioners building and operating Voice AI systems in real-world settings. The session follows an Ignite-style format with short, structured talks to enable clarity and comparison across layers of the stack. The focus is on what is working today, where constraints exist, and which decisions will matter over the next 12–18 months. The session concludes with a discussion that connects insights across the Voice AI stack for builders, adopters, and policymakers.

VIDEO RECORDING

Demystifying Voice Stack - What Makes Voice AI Work at Scale

Detailed Summary

  • The moderator explained the Ignite format: four speakers, each with seven slides and one minute per slide (≈ 7 minutes total). The focus would be on infra + telephony, evals + orchestration, policy + governance, and consumer/enterprise adoption. The goal was a mental model of the voice‑AI stack, trade‑offs, and actionable signals for builders, adopters, and policymakers.

2. Sunil Gupta – Infrastructure & Telephony Layer

Key PointsDetails & Nuance
Two Parallel Telecom Networks• PSTN (old‑generation circuit‑switched) network for feature‑phone/landline calls.
• Data‑network (IP‑based) for WhatsApp, Signal, Telegram‑style calls. Both must be supported at scale.
End‑to‑End Latency BottlenecksVoice AI must appear real‑time (sub‑second). Any perceptible delay (≥ 2–3 seconds) leads users to abort the call. Latency stems from (a) telecom transport, (b) data‑center processing, and (c) model inference.
Compute Sizing & Cost• Capacity planning must anticipate peak spikes (e.g., festivals, elections).
• Hybrid scaling: Base load on modest GPU fleet; burst capacity via on‑demand cloud contracts that can “burst” to high‑end GPU pools.
Redundancy & Fault‑ToleranceAll components (GPUs, storage, network links) must be active‑active to survive single‑point failures without customer impact. GPUs in production fail daily; systems must auto‑failover.
Edge‑AI ImperativeTo meet latency, edge data‑centers (regional compute nodes) should host the inference stack near the caller (e.g., a Guwahati‑based node for a Guwahati call). This mirrors the emergence of edge‑cloud in India.
Sovereign‑Cloud Considerations• Centralised public clouds raise data‑residency, compliance, and risk concerns (DPDP, RBI guidance).
• Sovereign cloud—operated within India’s legal jurisdiction—offers a middle ground, enabling both scalability and regulatory compliance.
Economic ScalingMassive scale will bring economies of scale (lower per‑minute cost). The speaker likened this to the UPI model where a service becomes effectively free due to indirect benefits.
AnnouncementsNo formal product launch, but a clear call‑to‑action: architect for active‑active, edge‑first GPU infrastructure and secure sovereign‑cloud contracts before large‑scale roll‑out.

Speaker Uncertainty – The transcript contains a few garbled phrases (e.g., “the voice air will work”) – interpreted as “voice AI will work”.

3. Maitreya Wagh – Orchestration & Multilingual Voice AI

Key PointsDetails & Nuance
Why India Needs a Dedicated Orchestration Layer• Privacy & Data‑Residency – Models must run on Indian sovereign clouds to protect sensitive data (healthcare, BFSI, government).
• Cost Sensitivity – Voice‑AI replaces cheap labor; even a 0.2‑rupee per minute saving matters.
• Multilingual Demands – Users mix code‑switching (Hindi‑English‑regional) within a single utterance; orchestration must detect language on‑the‑fly and select appropriate speech‑to‑text / LLM / text‑to‑speech models.
Complexities of Multilingual Voice AI1. Language Detection at Call Start – Identify the speaker’s preferred language within the first few seconds.
2. Hybrid Language Tokens – Users embed English terms (e.g., numbers, email addresses) within Hindi sentences; the system must respect token‑level language preferences.
3. Dynamic Prompt & Response Generation – Pre‑recorded system messages (e.g., “Fetching your calendar”) must be rendered in the current language, not in a fixed English fallback.
4. Document Translation on‑the‑fly – PDFs, web pages, etc., must be translated within < 100 ms to keep the conversation fluid.
Human‑in‑the‑Loop (HITL) Design• Graceful escalation – Detect when the AI can’t satisfy the request; transfer to a human agent without the user needing to request it.
• Quality Definition – Humans define what “good” means for each use‑case (e.g., “transfer to human after 2 failed attempts”).
Evaluation, Analytics & Audits• Continuous testing – Regression suites that simulate end‑to‑end calls across model combos.
• Observability – Granular logs for each stage (STT, LLM, TTS, telecom) to pinpoint failures.
• Audits – Periodic checks to ensure compliance with privacy regulations and internal quality standards.
Model Marketplace & Dynamic Selection• No single “best” model; each call may route to different STT/LLM/TTS providers based on cost, latency, and language quality (e.g., “11 Labs for modern English”, “Cartesia for Desi accent”, “Sarvam for cheap inference”).
• Orchestration UI functions as a dropdown to swap providers as they evolve.
Failure Modes1. Opaque error sources – Without end‑to‑end tracing, it’s unclear whether a failure came from STT, LLM, TTS, or telecom.
2. Cascading impact of prompt changes – Small prompt tweaks can break downstream components, requiring exhaustive regression testing.
AnnouncementsThe speaker did not announce a new product but highlighted the launch of Bolna’s orchestration platform that supports per‑call model selection and multilingual handling.

Speaker Uncertainty – The transcript contains a long garbled name sequence (“Mathrya Waag … 건…”)—interpreted as the introduction of Maitreya Wagh.

4. Deepika Mogilishetty – Policy, Governance & Data‑Protection

Key PointsDetails & Nuance
Voice as Biometric DataVoice recordings are personally identifiable information (PII). Consent is only the first layer of responsibility; additional safeguards are required to avoid misuse.
Regulatory Landscape (India)1. DPDP (Digital Personal Data Protection) Act – Enforced May 2025; mandates explicit consent, data‑minimisation, and auditability.
2. AI Governance Framework – Non‑binding guidance encouraging responsible AI development.
3. Sector‑specific guidelines (e.g., BFSI/RBI) – Additional controls for financial services.
Three‑Layer Overlap• National law (DPDP)
• AI‑specific framework
• Sector‑specific regulations – All must be satisfied simultaneously.
Consent vs. Trust• The “consent box” is insufficient; the real aim is to build a safe user experience that earns trust.
• Voice AI’s human‑like quality can increase user vulnerability (people share more sensitive info).
Liability & Accountability• Graded liability chain – Responsibility is distributed across data collectors, model providers, and platform operators.
• Audit trails – Record decision‑making, model versions, and data handling for compliance.
Practical Recommendations1. Map data flows – Visualise ingestion, processing, storage, and deletion pipelines.
2. Implement “right‑to‑delete” pipelines as rigorously as collection.
3. Enforce purpose‑limitation, minimal collection, and limited retention.
4. Localise processing – Keep data within Indian jurisdiction whenever possible.
5. Annual data audits – Demonstrate “duty of care” rather than merely avoiding penalties.
Policy InsightThe speaker urged the ecosystem to treat the AI governance framework as a law‑like requirement (similar to helmets or traffic rules) rather than optional guidance.
AnnouncementsNo product launch; emphasis on adoption of DPDP‑compliant practices across voice‑AI pipelines.

Speaker Uncertainty – Some sentences were truncated; the core message has been reconstructed from context.

5. Debdoot Mukherjee – Enterprise & Government Adoption Perspective

Key PointsDetails & Nuance
Why Voice AI Matters for India• Thinking‑vs‑Typing Gap – Users comfortable speaking regional languages struggle with Latin‑script typing; voice removes this friction.
• Discovery Challenge – Traditional feed‑based UI limits conversational discovery. Voice acts as a hands‑free overlay that does not disrupt UI real‑estate.
Trust Barriers1. Unclear expectations – Users often unsure whether the bot can solve their problem.
2. Perceived unnaturalness – Scripted or robotic speech erodes trust.
3. Multilingual & dialect diversity – 700+ active Indian dialects demand localized models; lack of dialect‑aware agents reduces user delight.
Technical Constraints• Background noise & low‑bandwidth networks – Must robustly handle noisy, low‑speed connections common in many Indian regions.
• Latency & Consistency – Calls that drop within 10 seconds are common; consistent problem‑solving is essential for retention.
Operational Recommendations1. Identify high‑volume, low‑risk use‑cases before tackling complex, high‑stakes scenarios.
2. Prioritise reliability & fault‑tolerance – A single failure can break trust.
3. Track leading indicators (e.g., repeat‑call rate, CSAT, CES) alongside business metrics.
4. Build deep observability & feedback loops – Continuous discovery of failure points and rapid iteration.
Metrics & ROI• A/B testing to measure conversion lift, CSAT, etc.
• Proxy metrics (e.g., “repeat call within 24 h”) as early signals of success.
• Cost‑saving – Voice AI reduces manpower, office overhead, and enables massive parallel outreach (e.g., 10 k simultaneous interview calls).
Panel Discussion – “One Unlock for the Next 12‑18 Months”• Sunil Gupta – Empathetic conversational capability (deep empathy, human‑like caring) as a research breakthrough.
• Maitreya Wagh – Mindset shift: treat consent & policy not as an after‑thought but as a core trust‑building pillar.
• Deepika Mogilishetty – Full compliance with DPDP and a graded liability model to foster accountability.
• Debdoot Mukherjee – Ensuring the AI openly declares its nature (“I am an AI”) to build trust, and guaranteeing ultra‑low latency and seamless hand‑off to humans.
AnnouncementsNo formal product announcements; the speaker called for industry‑wide adoption of best‑practice metrics and human‑centric design to unlock scale.

Speaker Uncertainty – Minor transcription glitches (e.g., “Santosh offline”) were interpreted as “Santosha offline” – likely a slip of the moderator’s name.

6. Closing Remarks

  • The moderator thanked the panelists and noted that the session provided the richest content of the summit. The discussion underscored that scaling voice‑AI in India requires tight integration of infrastructure, orchestration, policy compliance, and enterprise‑centric design. The session was concluded with applause and a brief thank‑you from the audience.

Key Takeaways

  • Infrastructure is the bottleneck – Latency, redundancy, and edge‑computing are non‑negotiable for real‑time voice‑AI at population scale.
  • Sovereign cloud offers a pragmatic bridge between scalability and Indian data‑residency/compliance requirements.
  • Orchestration must be multilingual and dynamic; language detection, code‑switch handling, and per‑call model selection are essential.
  • Human‑in‑the‑loop designs and granular observability prevent silent failures and enable rapid iterative improvement.
  • Voice data is biometric; consent is only the first step. Full DPDP compliance, audit trails, and graded liability are needed to earn user trust.
  • The “thinking‑vs‑typing” gap makes voice a uniquely powerful access channel for India’s diverse linguistic population.
  • Trust barriers stem from unclear expectations, unnatural speech, and dialect mismatches; addressing them requires empathetic design and clear AI disclosure.
  • Success metrics must blend business outcomes (conversion, CSAT) with leading indicators (repeat‑call rate, latency) and robust feedback loops.
  • One unlock for the next 12‑18 months (as voiced by the panel) is a mindset shift toward trust‑centric, compliant, and empathetic AI—combining technical breakthroughs (edge AI, empathy) with policy‑first thinking.
  • Edge‑first, sovereign‑cloud‑first architecture, coupled with multilingual orchestration and policy‑driven data governance, forms the roadmap for scaling voice‑AI across India.

See Also:

Graph View

Backlinks