AssemblyAI

Today we're shipping a major upgrade to streaming diarization, and it pulls us decisively ahead of the competition on the metrics that matter in production. Head-to-head vs. the competition: 🎯 2x better cpWER on 2-speaker telephony 📊 13% better cpWER on 4-speaker meetings 🔇 42% fewer false-alarm speakers 👻 91% fewer phantom turns and words attributed to speakers who don't exist For an AI notetaker, the 91% reduction in phantom-speaker words is the difference between a clean transcript and one your customers have to hand-correct. For an agent-assist tool, it's the difference between coaching prompts based on what the customer actually said and prompts generated from words the customer never spoke. We also updated the API: every word object now carries its own speaker label, unlocking mid-turn speaker change detection at the word boundary instead of the turn boundary. ✅ Live today. Learn more: lnkd.in/eCagsaia 👈

@jayy_1007 Excited to see what you put together!

Implementing voice agent API of @AssemblyAI in oakgen.ai !! Going to make it live soon, stay tuned ! btw @AssemblyAI did a good job with it, impressed with the output and pricing as well.

A voice agent. One prompt. Under 15 minutes. That's what Mart built using the AssemblyAI Voice Agent API and Claude Code—and we captured the whole thing on video. Here's what the build actually looked like: 🔹 Install the AssemblyAI MCP server → docs auto-inject into your Claude Code session 🔹 Drop one prompt describing your agent → Claude Code writes frontend and backend 🔹 Deploy to Railway → authenticate via backend token (no exposed API keys) 🔹 Add tool calling with Exa Search for source-backed responses 🔹 Let users pick from AssemblyAI's full voice library at session start If you've been sitting on a voice agent idea, this is the fastest path from concept to production we've seen. Watch the full build-along 👇 youtube.com/watch?v=E6AZhC… If you try it—drop your favorite voice in the comments. Our team wants to know. 🎙️

Introducing the Voice Agent API. One WebSocket. Stream audio in, get audio back. We handle the full voice stack so you can focus on your product. Powered by Universal-3 Pro, our speech model built for real-world audio. $4.50/hr. No SDK. Ship today → assemblyai.com/voice-agent

@agermanidis @runwayml @btaylor @SierraPlatform Dylan Fox, founder and CEO, @AssemblyAI. All talk. All action. May 6 in San Francisco. Apply now to join us: cerebralvalleyvoice.com

@gonzaloalfarof @OpenAI Glad you like it!! 🎙💙

if you think @OpenAI whisper is impressive, wait until you try @AssemblyAI's models, they’re on another level, try it and see the difference 🚀

@NeurlCreators Love this!! 💙🎙️

Typing commands to your AI agent? There's a better way. We integrated AssemblyAI's Universal-3 Pro into OpenClaw for voice interaction. Now you can control your agent through voice notes in Telegram. Custom keyterms, filler detection, acoustic cues, all handled automatically. Demo + code: neurlcreators.substack.com/p/building-a-v… #AIAgent #SpeechToText #OpenClaw #AssemblyAI

@odelbos Nice!! 🔥

Realtime Speech-to-Text using #AssemblyAI and Google translations. It's a simple #PoC using #Svelte for recreational programming. Code on my Github : github.com/odelbos/poc-sp… #realtime #speech_to_text #assembly_ai #google_translator #svelte #llm #ai

Vibe coding just leveled up. We brought voice mode to Claude Code using AssemblyAI's Universal-3 Pro Streaming. Why type your prompts when you can just say them? You get insane entity accuracy from AssemblyAI and the full power of Claude Code, all hands-free. Here's the full command: ASSEMBLYAI_API_KEY=[YOUR-API-KEY-HERE] bash -c "$(curl -fsSL assembly.ai/voice)" And get a free API key from your dashboard: assemblyai.com/dashboard/api-… Enjoy! 😎🎙️🎧

@YouveGotFox was on stage at @HumanXCo this week, and one thing he said captures how we think about building at AssemblyAI. "You always find new things once you go live." No matter how well you plan an AI deployment, the edge cases that actually break things are invisible until real users show up. The teams getting this right aren't the ones who anticipated every failure mode. They're the ones who built for visibility—good telemetry, tight feedback loops, and the ability to ship a fix fast. At AssemblyAI, this is how we approach building on every team. The gap between a struggling AI deployment and a successful one usually isn't the model. It's whether your team can see what's breaking and move quickly enough to do something about it. Glad to be at @HumanXCo with builders from around the globe!

Built with AssemblyAI! 🎙️💙

Medical Mode catches it before it gets that far. Works on both Pre-recorded and Streaming audio. HIPAA BAA included. $0.15/hr. See our benchmarks here → assemblyai.com/medical-mode Test with your own audio → assembly.ai/playground

The real failure mode isn't the transcript. It's what comes next. Most healthcare AI pipelines feed transcripts into an LLM → SOAP notes, discharge summaries, referral letters. Wrong drug name in. Wrong drug name out. Errors don't attenuate. They propagate.

General-purpose ASR: 95%+ accuracy on a clinical consult. Also general-purpose ASR: gets "hydrochlorothiazide" wrong every time. Introducing Medical Mode — a correction pass on top of Universal-3 Pro optimized for medical entity recognition. Enable it with one parameter.

Try Medical Mode today: assemblyai.com/introducing-me…

The Pitt meets AssemblyAI Medical mode 👀

Medical Mode is now available for clinical workflows. We built Medical Mode because a transcript that's 95% accurate can still be unusable in a clinical setting. Errors in general-purpose ASR are often concentrated on exactly the tokens clinicians care about most: drug names, dosages, and clinical terminology. "Lisprohumalog" is a phonetically reasonable guess. It's also not a real medication. Most healthcare AI products feed a transcript into an LLM to produce structured output. A wrong drug name in the transcript becomes a wrong drug name in the SOAP note, the discharge summary, the referral letter. Errors don't attenuate through the pipeline. They propagate. Medical Mode runs a correction pass optimized specifically for medical entity recognition: drug names, procedures, clinical terminology. The base model's noise handling and latency characteristics stay the same. Medical Mode just refines the output on the tokens that actually matter. Works on both Universal-3 Pro pre-recorded and Universal-3 Pro Streaming. No commitments or up-charges for BAAs to meet HIPAA compliance. 🔗 Try Medical Mode today: assemblyai.com/introducing-me…

🔗 Register for our workshop on March 31 assemblyai.zoom.us/webinar/regist…

🔗 Read the full blog post: assemblyai.com/blog/new-word-…

Most speech-to-text benchmarks are broken. Not because the tools are bad—because the truth files are. When we launched Universal-3 Pro, some customers flagged that their benchmarks showed the new model performing worse than older ones. So we dug in. What we found: the model was inserting words that weren't in the human truth files. When we listened back to the audio, the vast majority of those "errors" were words genuinely spoken—ones the human transcriptionist had missed. The better your AI gets, the more it exposes flaws in the ground truth it's being measured against. We built tooling to fix this—corrected truth file workflows, semantic word lists, and a GitHub repo to help you build benchmarks that hold up in production. You can test this tool on your own, or come learn how to use them on March 31 at our hands-on session on truth files, Semantic WER, and production-ready benchmarking.