Ce que Deepgram a publié en mai 2026

Deepgram upgrades Nova-3 Medical batch model with expanded medical vocabulary

Deepgram released an upgraded Nova-3 Medical batch model with expanded medical vocabulary and improved medical term recognition (97.20% KRR). The update maintains word error rate parity and is availab…

Speech-to-Speech vs Cascade: Voice Agent Architecture

Deepgram’s guide compares Cascade and Speech-to-Speech (S2S) voice agent architectures, emphasizing tradeoffs in cost, debuggability, and compliance. Cascade pipelines expose text at each stage, aidin…

Dynamic Range Compression for Voice AI

Deepgram’s product marketing manager argues that dynamic range compression (DRC) is often unnecessary for voice AI pipelines and can degrade transcription accuracy. The article provides a decision fra…

AI Voice Agents in Healthcare: 7 Production Use Cases (and What Makes Them Work)

Deepgram’s article highlights how health systems deploy AI voice agents for scheduling, refills, and triage, emphasizing the critical role of the speech-to-text (STT) layer in production success. It d…

Everybody's building Voice AI for restaurants right now. Let's draw the map.

The restaurant Voice AI market is rapidly saturating with developers, tech platforms, and enterprise brands building solutions. Deepgram positions itself as the foundational speech recognition layer e…

Deepgram Self-Hosted May 2026 release adds profanity filtering and Korean spacing fixes

Deepgram’s May 26, 2026 self-hosted release (260528) introduces profanity filtering for multilingual Nova-3 models and improves Korean word spacing in transcripts. The update also preps deployments fo…

Pricing change detected for Deepgram

Pricing updated for Deepgram: - Pay-As-You-Go: promotion changed — Free $200 Credit - New tier: Custom (Custom pricing) - Removed: Growth tier

Voice Agents That Prioritize Data Security and Run Where Your Data Lives

Deepgram launched a Voice Agent API that integrates NVIDIA Nemotron models (Nemotron 3 Nano and Super) to enable sub-700ms end-to-end latency for voice agents deployed in customer VPCs, on-prem, or hy…

Deepgram launches self-hosted voice AI deployment with enterprise prerequisites

Deepgram introduced self-hosted deployment options for its voice AI services, targeting use cases with strict performance or security requirements. The offering requires an Enterprise Plan, direct lic…

Pricing updated for Deepgram: - Pay-As-You-Go: promotion changed — $200 free credit then pay-as-you-go - New tier: Growth ($4000/yr ($333.33/mo annually, ~12.5% off)) - Removed: Enterprise tier

AI Voice Agents Improve Patient Engagement: What the Evidence Actually Shows

Deepgram’s analysis finds no peer-reviewed studies validating vendor claims of 30–50% appointment booking lifts from AI voice agents in healthcare. Most evidence comes from vendor case studies with sh…

Hinglish: The Language 600M+ Indians Speak and Why Your Voice AI Keeps Failing

Deepgram introduced multilingual code-switching capabilities in its speech-to-text API to handle Hinglish, a blend of Hindi and English spoken by 600M+ Indians. The feature detects language shifts wit…

Why Word Error Rate Is Broken for Indian Languages: The BRIDGE 7-Metric Stack Explained

Deepgram argues that Word Error Rate (WER) systematically overstates errors for Indian languages due to morphological agglutination, script diversity, and code-switching. They propose the BRIDGE 7-met…

Evaluating Voice AI Agents for Healthcare: The Compliance and Accuracy Checklist You're Missing

Deepgram released a detailed checklist for evaluating voice AI agents in healthcare, emphasizing the intersection of HIPAA compliance and transcription accuracy. The guide highlights medical-specific …

What is Code-Switching? A Complete Guide for ASR Builders

Code-switching in speech can cause ASR error rates to spike up to 11x higher, with monolingual systems failing at language boundaries. Unified multilingual models and specialized metrics like PIER are…

Deepgram adds Gemini 3.5 Flash to Voice Agent API and deprecates older models

Deepgram introduced the managed Google Gemini 3.5 Flash model in its Voice Agent API, replacing the older Gemini 2.5 Flash family. The new model improves performance and efficiency, while the 2.5 Flas…

How AI Voice Agents Work: A Beginner's Guide

21

Deepgram expanded profanity filtering to all multilingual models (Nova-2, Nova-3, Flux) via the profanity_filter=true API parameter, replacing inappropriate language with asterisks. It also fixed miss…

Deepgram adds Google’s Gemini 3.1 Flash Lite to Voice Agent API

Deepgram introduced the managed Google LLM model `gemini-3.1-flash-lite` in its Voice Agent API, replacing the preview version. The deprecated `gemini-3.1-flash-lite-preview` will be removed on May 26…

Call Center Compliance Regulations and Action Plan for 2026

Call center compliance in 2026 will require managing four regulatory layers: federal (TCPA, HIPAA, PCI DSS 4.0), state AI disclosure laws (Utah, California, Texas), international rules (GDPR, EU AI Ac…

Playground vs API: The Hidden Pronunciation Gap in Modern TTS

Deepgram highlights a critical gap where playground TTS demos mask pronunciation failures that emerge in production, particularly with raw user inputs like numbers, proper nouns, and domain terms. The…

Chatbot vs. Conversational AI: Key Differences Explained

Medical Voice Recognition: A Beginner's Guide

Deepgram released a comprehensive guide explaining how medical voice recognition works in clinical settings, including HIPAA compliance requirements, EHR integration challenges, and accuracy benchmark…

Deepgram vs Rev AI: Which Speech-to-Text API is Best for Developers?

Gemini deprecations

Google’s official deprecation page outlines the end-of-life timelines for numerous stable and preview models in the Gemini API, including Gemini 3, 2.5 Pro/Flash, 2.0, and others. Shutdown dates are t…

Deepgram adds Gemini 3.1 Flash Lite to Voice Agent API

Deepgram introduced the managed Google LLM model `gemini-3.1-flash-lite` in its Voice Agent API, replacing the preview version. The deprecated `gemini-3.1-flash-lite-preview` will be removed on May 26…

Resources

Deepgram introduced Flux Multilingual, the first conversational speech recognition model supporting multiple languages without routing. The model expands speech-to-text capabilities to Thai, Cantonese…

How Voice AI Works: From Sound Waves to Smart Conversations

Deepgram Aura-2 adds runtime speed and pronunciation controls

Deepgram’s Aura-2 TTS voices now support runtime speed (0.7x–1.5x) and pronunciation overrides via inline IPA notation in English and Spanish. These controls are available on batch and streaming endpo…

Deepgram adds Numerals support for Russian, Romanian, and Hebrew

Deepgram expanded its Numerals feature to three new languages—Russian, Romanian, and Hebrew—using monolingual models. The update allows spoken numbers to be converted to digits in transcripts via the …

Nova-3 Expands Speech-to-Text Support for Thai, Cantonese, Mandarin, and Indic Languages

Deepgram’s Nova-3 speech-to-text model now supports Thai, Cantonese, Mandarin (Simplified and Traditional), and improved accuracy for Bengali, Marathi, Tamil, Telugu, and Gujarati. These additions tar…

Jobcase Delivers Faster, More Natural Voice Experiences with Deepgram Aura-2

Jobcase adopted Deepgram’s Aura-2 text-to-speech to enhance its AI voice agents, reducing latency and improving naturalness in calls for job-seeking members. The integration supports both inbound and …

Klubi Transforms Brazil’s Interest-Free Group Purchasing Funnel and Scales Voice-Led Growth with Deepgram

Klubi, a Brazilian digital consórcio platform, scaled voice-led growth using Deepgram’s Nova-3 speech-to-text to automate pre-sales, qualification, and post-sales workflows. The integration enabled re…

Creditas Transforms Indian Debt Collections with Deepgram

Creditas, an Indian digital debt collections platform, adopted Deepgram’s speech-to-text API to automate and enhance collections while ensuring compliance and trust. The solution provided 100% call au…

Abby Connect scales high-touch service and launches AI receptionist with Deepgram

Abby Connect, a 24/7 human virtual receptionist service, integrated Deepgram’s speech-to-text API to power its new AI receptionist, automating repetitive tasks like scheduling and FAQs. The integratio…

How Vida Delivers Empathetic Healthcare Voice Agents with Deepgram Aura-2

Vida, an AI Agent OS for enterprises, selected Deepgram’s Aura-2 TTS and multilingual STT to power high-volume healthcare voice agents, citing superior naturalness, low latency, cost predictability, a…

Leading Medical Tech Transcription Platform

Leading Cloud Communications Platform

A global cloud communications platform serving 19,000+ businesses replaced its in-house Whisper-based transcription with Deepgram’s AI speech-to-text platform on AWS. The move addressed scalability, a…

SigmaMind AI Powers a Million Monthly Voice Agent Calls with Deepgram’s Real-Time Speech-to-Text

SigmaMind AI integrated Deepgram’s Nova-3 and Flux speech-to-text models to power its no-code voice AI platform, reducing end-to-end agent response latency by 300ms and enabling mid-utterance API call…

How a Large Fortune 50 U.S. Retail Pharmacy Chain Scaled Automated Voice Solutions to 1M+ Calls Per Day with Deepgram

A Fortune 50 U.S. retail pharmacy chain replaced its legacy Nuance IVR with Deepgram’s Nova Medical speech-to-text and Aura text-to-speech to handle over 1 million pharmacy calls daily across 7,000+ l…

GetVocal scales governed enterprise voice automation with Deepgram

GetVocal AI integrated Deepgram’s real-time streaming speech-to-text into its voice automation platform to support production-grade telephony interactions. The integration improved structured entity c…

14

Deepgram launched profanity filtering for over 50 languages, enabling automatic detection and redaction of offensive language in transcripts via a simple API parameter. The feature targets cleaner, sa…

Deepgram Self-Hosted May 2026: Batch Diarization v2 now available

Deepgram’s May 14, 2026 self-hosted release (260514) introduces Batch Diarization v2, a significantly improved speaker-labeling model for pre-recorded audio. New deployments default to v2, while exist…

Put a Deepgram Voice Agent on Any Web App in Minutes

Deepgram introduced the Browser Agent SDK, four composable npm packages enabling rapid integration of voice agents into any web app. The SDK abstracts complex audio handling, reconnection logic, and s…

Your restaurant needs to speak Spanish, y ahora puede.

Deepgram introduced Flux Multilingual, a Voice AI model enabling real-time code-switching across 10 languages for restaurant ordering. The feature maintains monolingual-grade accuracy and latency whil…

Maki and Deepgram partner to power real-time voice AI for hiring

Maki partnered with Deepgram to embed real-time streaming speech-to-text at the core of its voice pipeline, enabling AI hiring agents to conduct natural, responsive candidate conversations with high a…

Smarter, Faster Calls for Every Business: Lindy Gaia Launches with Deepgram Flux

Deepgram launched Flux, a conversational speech recognition model, and partnered with Lindy to power Gaia, a no-code AI voice agent for handling business calls. Flux enables ultra-low latency and natu…

Deepgram enhances Nova-3 Portuguese model with improved accuracy

Deepgram updated its Nova-3 Portuguese model to improve transcription accuracy for both Brazilian and European Portuguese variants. Users can now leverage the enhanced model by specifying `model="nova…

Deepgram vs Amazon Transcribe: Which Should Power Your Voice App?

Deepgram published a detailed 2026 comparison of its speech-to-text service against Amazon Transcribe, highlighting differences in accuracy benchmarks, streaming latency, custom vocabulary features, p…

Deepgram vs Twilio: Key Differences for Real-Time Transcription

Deepgram explains when to use its direct API versus Twilio’s managed paths (Gather, ConversationRelay) for real-time transcription. Direct API access unlocks full STT control, Keyterm Prompting, and m…

On-Premise Speech-to-Text: Which STT API Offers True Data Control?

A comparison of self-hosted speech-to-text (STT) options from Deepgram, Speechmatics, AssemblyAI, AWS, and Google Cloud highlights varying levels of data control, air-gap support, and compliance certi…

AI Drive-Thru: How Voice AI Is Transforming Order Taking

Deepgram introduced AI drive-thru ordering using its Nova-3 speech-to-text model, trained on real drive-thru audio, achieving a 5.26% word error rate. The system integrates with POS in real time and u…

Deepgram releases SDK updates with Flux multilingual support and breaking changes

Deepgram shipped SDK updates across JavaScript, Rust, Python, and Java, adding Flux multilingual support in Rust, restoring the Agent interface in JavaScript, fixing WebSocket query parameters in Pyth…

Deepgram launches composable Browser Agent SDK with four packages

Deepgram introduced a new Browser Agent SDK with four composable packages—Widget, React UI Components, React Hooks, and JavaScript SDK—enabling quick integration of voice agents into web apps. Each pa…

Deepgram launches Browser Agent SDK with composable packages for Voice Agent API

Deepgram introduced the Browser Agent SDK, offering four composable packages to connect web apps to the Voice Agent API, including a drop-in widget and React components. The SDK simplifies integration…

Deepgram released a Browser Agent SDK with four composable packages—@deepgram/agents-widget, @deepgram/ui, @deepgram/react, and @deepgram/agents—enabling web apps to connect to the Voice Agent API. Th

Pricing updated for Deepgram: - New tier: Pay-As-You-Go (Custom pricing) - New tier: Enterprise (Custom pricing) - Removed: Free tier - Removed: Custom tier

A margin doomspiral is happening in QSR, and Voice AI is here to save your P&L

Deepgram introduced Deepgram for Restaurants, a Voice AI system designed to address the QSR industry's labor-driven margin crisis. Rising wages and 130% annual turnover are crippling QSR operators, le…

Build Voice Agents in Your AI Coding Tool

Deepgram introduced three agentic engineering tools—the dg CLI, MCP server, and deepgram/skills repo—to streamline voice AI development in AI coding tools like Claude Code and Cursor. These tools auto…

Virtual Event: How to Evaluate TTS for Voice Agents | Beyond the Vibe Check

Deepgram’s TTS team will host a live webinar on May 5, 2026, to teach engineers and product teams how to build scalable TTS evaluation pipelines for voice agents. The session covers defining scoring c…

Best Voice AI Platforms for Banking in 2026

The article provides a framework for evaluating voice AI platforms in banking, emphasizing accuracy under real-world noise, compliance architecture, and cost predictability at scale. It highlights Dee…

AI Voice Agents for Business: A Buyer's Guide

Deepgram published a comprehensive buyer's guide comparing AI voice agent services based on latency, noise tolerance, pricing, and compliance. The guide emphasizes testing under real-world conditions,…

Best Vapi AI Alternatives for Voice Apps

A third-party guide ranks Vapi alternatives for voice applications based on STT accuracy, pricing transparency, and deployment flexibility. Deepgram is highlighted for production-grade STT and bundled…