Deepgram

🔊Tomorrow: Deepgram’s TTS product and research teams walk through how to build a TTS evaluation pipeline for production voice agents. Scoring rubrics, automated metrics, structured human eval. 10am PT / 1pm ET. Register Now: deepgram.com/webinars/how-t…

🌐 “Multilingual” shouldn’t mean multi-system. Yet most real-time speech stacks still look like this: → language detection service → per-language models → routing logic → latency you can’t explain There’s a better way. 📈 Deepgram’s Flux Multilingual runs 10 languages in a single real-time stream — with: ✅ automatic language detection ✅ native code-switching ✅ language_hint biasing ✅ one model, one WebSocket No routing layer. No model hopping. No duct tape. If you’re building voice agents, this changes your architecture. Read the technical deep dive ↓ deepgram.com/learn/flux-mul…

Full-stack Voice AI in one pipeline? Yes—here’s what it actually looks like ⚙️🔥 This deep dive by @asifrazzaq1988 shows how to build an end-to-end workflow with Deepgram’s Python SDK: real-time transcription, async audio processing, TTS generation, and text intelligence (sentiment, topics, intents). We’re talking word-level timestamps, diarization, summaries, keyword boosting, and scalable async execution. If you’re building Voice AI systems, don’t skip this 👇 marktechpost.com/2026/04/24/a-c…

@Cloudflare Workers AI's speech layer runs Deepgram Nova-3, Aura-2, and Flux at 300+ edge locations worldwide. For developers, that means voice agents that answer in the time it takes a caller to finish breathing in. Cloudflare handles TLS, DDoS, and caching; Deepgram handles the speech. Cloudflare is the platform, the voice models are powered by Deepgram. Build on Cloudflare Workers AI with Deepgram: developers.cloudflare.com/ai-gateway/usa…

Flux Multilingual is live. Real-time conversational speech-to-text for voice agents in 10 languages, with monolingual-grade accuracy, turn detection, and code-switching. Deploy once and launch globally. Learn more → deepgram.com/learn/introduc…

🔥 Hot take: voice agents aren’t chatbots with a microphone. They’re real-time distributed systems that have to coordinate perception, reasoning, timing, and response - all while a human is mid-sentence and your latency budget is measured in milliseconds. We put together a 107-page guide covering how to actually build these systems for production. Architecture, deployment, compliance, operational patterns - the works. 📝 Built by the team behind 200,000+ developers. Zero fluff. Link: deepgram.com/the-definitive…

"[Y]ou can now build an agent that users can talk to in real time over a single WebSocket connection, while keeping the same Agent class, Durable Object instance, and the same SQLite-backed conversation history. " - @Cloudflare `@cloudflare/voice` is an experimental package for the Agents SDK that provides many incredible features, including built-in Workers AI providers, so that you can get started without external API keys: ✅ Continuous STT with Deepgram Flux ✅ Continuous STT with Deepgram Nova 3 ✅ Text-to-speech with Deepgram Aura Learn more at this blog: blog.cloudflare.com/voice-agents/

🔊 Most voice agent teams evaluate TTS by listening and deciding if it sounds “good enough.” That works early on. It breaks down when you need to make real decisions across models, voices, and use cases. 📊 On May 5, Deepgram’s TTS product and research teams walk through how to build an evaluation pipeline: scoring criteria, automated metrics, and structured human eval. Register here: deepgram.com/webinars/how-t…

Wrapped up at the Restaurant Leadership Conference last night. One pattern across nearly every conversation with restaurant CTOs and engineering leaders: They're not shopping for a drive-thru plug-in. They want a platform their engineers can build on. We’re focused on solving the thorniest problems in voice AI so our partners can focus on delivering best-in-class operations: ✅ Sub-200ms latency at scale ✅ Speaker separation in a crowded drive-thru ✅ Scripting and menus that change every week ✅ Noise reduction ✅ Echo cancellation ✅ Accent and dialect robustness ✅ Turn detection and other models fine-tuned for restaurant environments These are research problems. Working with Deepgram means getting access to the research teams solving them. We're the tide that lifts all ships – every breakthrough our teams achieve goes straight into the platform our customers build on. That's what we're doing at Deepgram for Restaurants: building for builders.

🥊 We put OpenAI Whisper head-to-head with Deepgram Nova-3 across real-world audio, complex terminology, and edge-case scenarios that actually matter. The result? A bigger accuracy gap than most expect, especially for domain-specific terms, jargon, and specialized vocabulary. 📊 If you’re building Voice AI or deploying voice agents, transcription quality is foundational. Watch the full breakdown and see the data for yourself: youtu.be/l8Zj_iAnnd0

Thousands of restaurants rely on Slang AI to keep seats full every night and run their business. Restaurants that don’t use AI-enabled phone systems miss on average 150 calls a month, which is roughly $28K in lost revenue per year, per location. When a caller doesn't feel heard, they usually hang up and call somewhere else. But Voice AI needs to do more than just transcribe. It needs to actually understand people, accurately and reliably, in restaurant-trained environments that are noisy and fast-paced. @slang_ai migrated to Deepgram to achieve 43% fewer errors and 72% faster responses than Google¹ -- a noticeable difference for both Slang AI and its customers. Slang AI, powered by Deepgram. "Slang AI answers every call for thousands of restaurants. That only works if the Voice AI behind it is fast and accurate. Deepgram gave us a measurable improvement, and our restaurants notice the difference.” — Urim Choi, VP Product & Technology, Slang AI #VoiceAI #DeepgramForRestaurants #RestaurantTech ¹ - Pipecat's open-source STT benchmark (1,000 samples, 2026)

🇬🇧 London, we'll see you at @AWS Summit on 22 April! Deepgram is at booth B20 all day. Our engineers will be there running live demos and answering the voice AI questions you can't get a straight answer to over email: Speech-to-text, text-to-speech, voice agents. Anything you're building. Come say hi! 😄

🚀 Next week, the Deepgram for Restaurants team is headed to the Restaurant Leadership Conference! This is the room we've been wanting to be in. 1,700+ industry leaders all thinking about what's next for restaurants. We may be biased, but we think it has everything to do with voice AI. 🔊 If you've been wondering how AI can actually help the chaos of real restaurant operations, that's kind of our whole thing. Reach out and let’s grab a coffee (or a donut)! ☕ 🍩

Calling all Michigan builders 🦞 Deepgram is headed to ClawCon Michigan this Thursday April 16th at the Crisler Center, and we’re excited to meet the students, founders, and developers shaping the future of AI. 🌐 Come connect with our team, talk Voice AI, and share what you’re building! Details: Ann Arbor | April 16 | 6–9 PM ET See you at ClawCon.

"Deepgram, currently valued at $1.3 billion, has achieved adoption among many companies,... with a platform that can deploy voice agents. The company’s technology is also used to power over 2,000 third-party voice products." - ​Tim Keary, ​@Forbes ​​Check out the article linked below to hear more about our internal development practices, our partnership with IBM, and some words from our CEO @deepgramscott. Link to article: forbes.com/sites/timkeary…

🏆 Accuracy isn’t optional; it’s the foundation of speech-to-text AI. When people rely on your product, every customer interaction counts. Generic, one-size-fits-all AI models struggle, especially when your use case becomes increasingly specific. On the other hand, tailored speech-to-text and Voice AI like Deepgram's deliver the accuracy you need 🎯 Try it yourself → playground.deepgram.com #VoiceAI #SpeechToText #AI #CustomerExperience #AccuracyMatters

Something you may not expect as you drive up to a beloved nationwide drive-thru restaurant chain: the data that sits behind that light-up menu board is chaos. Duplicate items with different names. Inconsistent data between locations. Item IDs that read like OpenClaw agents spoke to each other a few too many turns. In Part 2 of our engineering culture series, we dig into the problem that nearly became our biggest scalability blocker: menu data. POS exports are a mess, and an AI needs to understand what a customer means when they say "I'll take a crispy chicken sandwich." One of our engineers saw the problem, opened a repo, and started shipping. What came out is an agentic menu integration pipeline that cuts menu onboarding from weeks to a day. Read the full post here: deepgram.com/learn/deepgram… If you're an engineer who wants to ship like this, we're hiring: deepgram.com/careers If you're a platform or restaurant tech company tired of hand-translating POS data for every integration, reach out here: deepgram.com/contact-us-res… #DeepgramForRestaurants #RestaurantTech #VoiceAI #Engineering

🏆 “If accuracy actually matters, custom models win.” In a world where people expect seamless experiences, your customer interactions can’t afford missteps. Yet too often, generic AI models fall short—especially when it comes to speech-to-text in real-world environments. That’s where voice AI built for customization changes everything. With Deepgram’s keyterm prompting, you get the accuracy, adaptability, and control needed to handle domain-specific language, complex workflows, and high-stakes conversations—without compromise. See the difference for yourself → playground.deepgram.com #VoiceAI #SpeechToText #AI #CustomerExperience #AccuracyMatters #ConversationalAI #DeveloperTools #EnterpriseAI

Tomorrow morning: Skip the line. Order with your voice. 🗣️ The Deepgram Voice AI Espresso Bar opens tomorrow morning at Booth #206! Experience our "drive-thru" in action—order your customized coffee directly from our AI agent and see just how fast real-time voice can be. 👇 See you at the booth tomorrow morning!