Descriptions:
Ashwyn Sharma, CEO and co-founder of Vocal Bridge—a company backed by Andrew Ng’s AI Fund and DeepLearning.AI—demonstrates how to add production-grade voice interfaces to existing applications and AI agents, drawing on a decade of shipping voice AI at Meta, Salesforce, and Audible. Presented at AI Dev 26 in San Francisco, the talk addresses a problem familiar to anyone who has attempted to ship voice features: what starts as a straightforward speech-to-text plus LLM plus text-to-speech pipeline quickly explodes into a months-long infrastructure project involving voice activity detection, endpointing, barge handling, and turn detection.
Sharp presents three integration surfaces that Vocal Bridge exposes. The first is adding voice to existing React applications using an npm-installable SDK, a VocalBridge provider wrapper, and two React hooks—onAction and sendAction—that create bidirectional communication between a voice agent and the application state. A live tic-tac-toe demo shows the voice agent maintaining full awareness of the game board, responding to both voice commands and mouse interactions in real time. The second surface is adding a voice layer to an existing text-based agent with a few lines of code; the demo uses a Claude-backed agent that handles domain queries while the Vocal Bridge layer manages small talk and modality switching. The third surface, voice as a tool, lets autonomous agents choose when to invoke voice modality based on context.
The talk includes a pip-installable CLI for configuring prompts and client actions, and walks through the token-minting API key flow for authentication. Engineers evaluating voice SDKs or looking to add voice capability to Claude-backed agents will find the implementation walkthrough directly actionable.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Hands On Build
