Faster ASR is not only a benchmark win.
It changes what teams can afford to build, from live captions to large-scale transcription to voice agents that feel responsive rather than procedural.
Read the breakdown: https://t.co/r7z2YK16TK
Nvidia has made ASR faster for a reason that should worry the rest of the market. βοΈπ§
The interesting move is architectural: Parakeet uses Token-and-Duration Transducers to predict what was said and how long it lasted.
That changes throughput.
For batch transcription/ pre-recorded audio, the better question is: what happens after the audio is processed? π€
And, can the model handle long files, accents, technical language, speaker changes, and downstream AI tasks without compounding errors?
Read our guide: https://t.co/sOniBB2mEI
Speech AI has a speed problem.
Not because it is too slow, but because too many teams measure speed like it tells the whole story. β±οΈπ
RTF is useful, but it will not tell you whether a transcript is accurate, readable, scalable, or fit for enterprise workflows.
The best testing platforms now look at the full interaction loop: speech recognition, reasoning, turn-taking, recovery, compliance, and observability.
That is where voice AI becomes measurable.
Read the guide: https://t.co/HDyjZTyC3h
Most voice agents sound impressive in a demo.
Production is where the illusion starts to sweat. ποΈπ§ͺ
The real test is not whether an agent can answer. It is whether it can cope with latency, noise, interruptions, accents, and users who do not behave like test scripts.
Is speech recognition really a π΄π°ππ·π¦π₯ problem β or have we just scratched the surface? Listen to this Convo AI World Podcast episode with @Speechmatics ' Ricardo Herreros Symons.
A summer thank you to everyone building with Speechmatics...
We're boosting the free tier for the next few months. π
More room to evaluate and prototype, on us. π
What's new:
π 20 hrs/month batch
π 20 hrs/month real-time.
Live now, applied automatically to every account.
Enough to run a proper eval on your own audio, prototype a captioning feature end-to-end, or put something working in front of real users. π
Amsterdam, Speechmatics aan de lijn! βοΈ
Speechmatics will be showcasing our Medical Model which is production ready across nine languages, including Dutch, German, and every Nordic language, with regional pharmaceutical naming, compound words, and clinical terminology built in.
Itβs the same engine thatβs helped partners like @Thymia build voice biomarkers
Stop by our booth for a post lunch sweet treat and see how it sounds in a real consultation at HLTH Europe.
β° 2:30 pm
π 16th & 17th June
πBooth D63
So we rebuilt for the machine in front of the editor: macOS, Windows, laptop GPUs, tight memory, and no room to slow Premiere down.
Read Part One of the Adobe story: https://t.co/AaCsa2jYPa
Adobe Premiere needed cloud-grade transcription without the cloud.
That sounds simple until you remember audio doesnβt behave like images or text. π¬
Voice AI has to handle messy, time-based data while staying fast enough for real creative workflows. π
Hospitals are loud. Demo rooms arenβt. π₯
Most speech recognition models still optimize for the demo room: clean audio, one speaker, no overlap, no accent. Real clinical audio is the opposite of that.
Our Speechmatics Medical Model is trained on the real version. Overlap, accents, fast turn-taking, equipment noise, code-switching. It hits 93% real-world accuracy, 96% medical keyword recall, and 50% fewer errors on medical terms than the next closest provider.
Your #1 choice for medical-grade voice AI. See it run live at #HLTHEurope, booth D63.
Book a meeting: https://t.co/tw5rDTRm03
In London for Tech Week? We're hosting a meet-up with @Agora and you're invited.
If you want to hear the latest and greatest in Voice AI and get a front row seat to panels with the greatest minds in Voice AI you wont want to miss this.
π£οΈ Builders, operators, and investors in one room.
π£οΈ Panelists from Mux, Neuphonic, and Inkling.
πTue 10 June at the Dickens Inn π»
Lots of food and drink to go around as always
Sign up here π
https://t.co/PFDRPtGB7X
#VoiceAI #LondonAISummit
One week out for the Voice AI Mixer.
πJune 10,London
Co-hosted by @Speechmatics, @AgoraIO and thymia at the Dickens Inn during AI Summit week.
Panels. Demos. Networking.
Feat. @RiquiHerreros (Speechmatics), Phil Cluff (@MuxHQ), Sohaib Ahmad (@neuphonicspeech), George Greenbury (Inkling AI).
Free to attend. A few spots left.
RSVP: https://t.co/PFDRPtGB7X
No infinite cloud compute. No perfect hardware. No room for lag when editors hit play.
In our new blog by Chief Architect Andrew Innes, part One tells the story of shrinking a cloud speech model for Premiere.
Read below + stay tuned for part two: rebuilding the on-device engine to take on Whisper. β‘
https://t.co/fyE3mAXUWW
Audio doesnβt play by the same rules as images or text.
That matters when your Voice AI has to run inside Adobe Premiere, not in the cloud. π¬
For Speechmatics, the challenge was making cloud-grade transcription work locally, across millions of laptops.
Prizes from @FeatherlessAI:
1st - Olympusos
2nd - PlutusAudit AI
3rd - Sentinel: Adversarial AI Court for Bodycam
@Speechmatics Prizes:
1st - Deals Machine: The Autonomous Sales Agent by View 1 Studio
2nd - Olympusos
3rd - Exam Dragon by Children of the Milky Way