Speechmatics

BREAKING 🚨 Pipecat voice agent evals are in. Highest accuracy on the board. 495ms median TTFS. Speed you can trust. 🧵

Fix this using Pipecat's interruption handling + Speechmatics STT to handle: 🔊Local audio capture 🔊VAD 🔊Speaker diarization 🔊Passive filtering 🔊Interruptions Find working code 👉 github.com/speechmatics/s…

Your voice agent needs to handle interruptions. e.g: Agent: ‘Hi, how can I help you?’ User: 'Can you tell me about—' Agent: keeps talking User: 'Actually… never mind' In fact, interruptions are the number 1 Voice AI frustration for users.

And try it out yourself via our Academy: github.com/speechmatics/s…

Want to differentiate speaker? 🗣️ Got a caller in New York and one in London - you can use multichannel diarisation that will have 100% accuracy. 👏 Take a look: youtube.com/watch?v=C3twU-…

What can a founder who switched from art to AI bring to voice? 🎨 We spent 5 minutes with @nimrod_ron to find out why he left painting for AI, what he'd open source tomorrow, and who he wants narrating when chaos hits. open.substack.com/pub/resonancev…

Join us + Voice AI spcce (@thetoolists) for: 🔊 Two tracks: Voice & Medical and Voice & Productivity 🔊 60 hand-picked builders 🔊 £1,500 best overall + £500 per track winner 🔊 Speechmatics API credits as an additional prize 🔊 Our team on-site: talk + mentorship throughout the day Well worth your Saturday. 📍Saturday 18th April // 10:00am – 9:00pm, London luma.com/voiceaihack

Healthcare is one of the most compelling spaces in Voice AI. The problems are structural: clinical notes that miss what was said, transcription that can't handle real-world conditions, patients navigating systems that weren't designed for them. The gap between what the tech can do and what's been built is still wide open. That's the brief for the Voice AI Hackathon in London on 18 April, and why we're backing it. 👇

POV: You work at @Speechmatics HQ ☀️👇 Today's calls will be taken from our rooftop 😎.

@pipecat_ai Head to our Academy to find working AI receptionist use cases: github.com/speechmatics/s…

Your AI receptionist shouldn't transcribe your Traitors episode playing in the background. 🧙 Speaker Focus via @pipecat_ai locks onto specific speakers 👇 Filter out everything else. Office chatter, background noise, the TV in the waiting room.

@NABShow Book a slot 👉 calendly.com/speechmatics-e…

Real-time STT. Sub-250ms. 55+ languages. On-device available. If you're building for broadcast, come see how @Speechmatics drowns out the noise. 🗓️ April 19–22. 📍 @NABShow, Booth N1304, UK Pavilion.

36 years in speech tech teaches you one thing fast: Voice AI works in demos. Production is where it gets real. 😬 Dr Tony Robinson went deep on Reddit last week with hard-won lessons for builders. Take a look 👇 speechmatics.com/company/articl…

Tony's full AMA. Essential reading for anyone building in Voice AI today 👇 speechmatics.com/company/articl…

Most Voice AI builders are blaming the wrong part of the stack. According to Dr Tony Robinson: It’s not usually ASR. It’s latency, LLM reliability, turn detection, and timing. We pulled some of his insights from a 24-hour Reddit AMA. Worth your scroll 👇

@bnicholehopkins also covers: 🔊 Why 90% accuracy will break your product 🔊 The testing trap most engineering teams walk straight into 🔊 Where voice architectures are heading by end of 2026 Full piece on Resonance: resonancevoiceai.substack.com/p/what-self-dr…

Self-driving cars solved a problem that voice AI hasn't figured out yet 🚗🎙️ @bnicholehopkins built eval infrastructure at Waymo. When voice agents started scaling, she saw the same mistake happening all over again. In this clip she explains exactly what she learnt, and why the industry is about to repeat the errors the autonomous vehicle world already made. 👇

🤝 New: test your STT against real production audio. We've partnered with @cekuraAi to bring head-to-head comparisons with Azure, Gemini, and Deepgram, on your audio, not benchmark datasets. All of this + more in Issue #9 👇 linkedin.com/pulse/issue-9-…

🌍 Arabic-English bilingual medical is fully live. Gulf, Egyptian, Levantine. Code-switching mid-sentence, no routing required. 35% fewer errors than the nearest competitor. 6.3% WER on mixed-speech. Realtime SaaS, Batch SaaS, on-prem. Learn more: speechmatics.com/company/articl…

dev🚨 /speechmatics Issue #9 is out. Feat. some big news for file transcriptions. A lot shipped this month. 👇