Madhav Lavakare retweeted
🏆 A new leader on the Voice Arena Diarization + ASR board: AirCaps A5S Async v1 joins the 14 systems already scored and enters at rank #1.
To put this result in context, we evaluated all 15 systems on the same 139 held-out conversations with 2–8 speakers, counting overlapping speech throughout. cpWER measures how accurately a transcript captures words and assigns them to the right speakers.
AirCaps A5S Async v1 scores 23.89% cpWER overall and leads in both settings, with 24.91% cpWER in-person and 17.47% in online calls.
What's particularly interesting is where this lead comes from. AirCaps and Nemotron 3.5 Transcribe (#2) miss a similar share of words (9.29% vs. 8.32%), but AirCaps has a substantially lower substitution rate (10.88% vs. 18.43%) and fewer insertions (3.72% vs. 5.68%).
AirCaps maintains its lead when we resample the 139 conversations. Its overall cpWER showing a 95% confidence interval of 19.47%–28.58%.
Congratulations to the @tryaircaps team on the launch! 👏🏻
Madhav Lavakare retweeted
Blows my mind what a tiny team of AI-pilled engineers can pull off! 🤯
@tryaircaps’ A5S Async v1 ranks #1 on VoiceArena for figuring out who said what. Across their five public benchmarks, the next-best model makes 41% more errors. They’ve open-sourced the benchmark so you can check the results yourself.
Small team. Big swing. Incredible work, @madlavak!
cc @vkhosla @khoslaventures
Introducing A5S Async V1: the world's most accurate speech recognition model for speaker-attributed transcription.
#1 on VoiceArena, ahead of ElevenLabs, AssemblyAI, Deepgram, NVIDIA, Meta & more.
The next-best model makes 41% more errors across 5 public benchmarks.
And yes, every lab says they're #1. So we open-sourced our benchmark.
Try it now👇
Really excited about this model. The gap between us and previously SOTA STT models like ElevenLabs, Nvidia, AssemblyAI, Google, Meta is surprisingly large. Proud of our tiny team taking on the titans. Try it out on the AirCaps API, first 5h free.
Introducing A5S Async V1: the world's most accurate speech recognition model for speaker-attributed transcription.
#1 on VoiceArena, ahead of ElevenLabs, AssemblyAI, Deepgram, NVIDIA, Meta & more.
The next-best model makes 41% more errors across 5 public benchmarks.
And yes, every lab says they're #1. So we open-sourced our benchmark.
Try it now👇
Built the world's most accurate speaker diarization model.
Beats ElevenLabs, AssemblyAI, Meta, Grok, Nvidia Nemotron, Pyannote, and others.
61-71% fewer errors than cloud ASR providers, 2.5x fewer errors than ElevenLabs Scribe V2. Benchmarks & methodology open-sourced.
Now your voice AI can know who said what!
Introducing AirCaps Diarization-v1: the world's most accurate speaker diarization model.
Beats every leading model across 7 popular diarization benchmarks.
61–71% fewer errors than leading cloud ASR providers with built-in diarization (ElevenLabs, AssemblyAI, Meta and Grok). 2.5x fewer than ElevenLabs Scribe v2, the best of them.
12.03% DER vs. 13.21% for DiariZen Large v2 and 13.44% for pyannoteAI Precision-3.
Built for distant mics, noisy rooms and overlapping speech.
Benchmarks, evaluation code & methodology open-sourced. Examples & access:
research.aircaps.com/diariza…
2
7
205
Madhav Lavakare retweeted
Insane that more people aren't talking about this! There's still a real gap right now for transcription in real world situations
Introducing A5S V2: SOTA streaming speech-to-text for the real world.
Ranks ahead of eight leading proprietary cloud streaming speech models in a real-world accuracy evaluation derived from four public noisy, multi-talker datasets.
A5S V2 achieves a 26.27% word error rate across the fixed four-corpus evaluation, beating ElevenLabs, Meta, AssemblyAI, OpenAI, xAI, Deepgram, and Google. The closest cloud model, ElevenLabs Scribe v2 Realtime, scored 26.64% WER.
36.7% lower WER than Nemotron. 47.9% lower than Whisper Streaming.
Runs locally on iPhone & MacBook, with 50 ms / 44 ms mean finalization latency in our device trials.
Built for noisy rooms, distant microphones, and overlapping speech.
Demos, benchmarks & evaluation access:
research.aircaps.com/a5sv2
Just released A5S V2: a SOTA streaming speech to text model built for the real world.
We started building a wearable and ended up building the speech model we needed.
Today, it runs on an iPhone and outperforms every cloud model we evaluated on hard real-world audio.
Introducing A5S V2: SOTA streaming speech-to-text for the real world.
Ranks ahead of eight leading proprietary cloud streaming speech models in a real-world accuracy evaluation derived from four public noisy, multi-talker datasets.
A5S V2 achieves a 26.27% word error rate across the fixed four-corpus evaluation, beating ElevenLabs, Meta, AssemblyAI, OpenAI, xAI, Deepgram, and Google. The closest cloud model, ElevenLabs Scribe v2 Realtime, scored 26.64% WER.
36.7% lower WER than Nemotron. 47.9% lower than Whisper Streaming.
Runs locally on iPhone & MacBook, with 50 ms / 44 ms mean finalization latency in our device trials.
Built for noisy rooms, distant microphones, and overlapping speech.
Demos, benchmarks & evaluation access:
research.aircaps.com/a5sv2
4
5
183
Team cooked w this one
Today, we’re publicly launching the AirCaps Audio Research Lab.
Real-world audio is 10x harder to understand than the digital world, and voice is far from “solved.”
We’re building fully streaming speech and audio models for the physical world: target speaker extraction, far-field enhancement, diarization, separation, and speech-to-text.
Our models are a step function ahead of leading cloud speech models in the hardest acoustic environments, while running on edge on consumer hardware.
See real demos: research.aircaps.com/
2
6
192
Madhav Lavakare retweeted
Today, we’re publicly launching the AirCaps Audio Research Lab.
Real-world audio is 10x harder to understand than the digital world, and voice is far from “solved.”
We’re building fully streaming speech and audio models for the physical world: target speaker extraction, far-field enhancement, diarization, separation, and speech-to-text.
Our models are a step function ahead of leading cloud speech models in the hardest acoustic environments, while running on edge on consumer hardware.
See real demos: research.aircaps.com/
I love YC. Grew monthly revenue 6.5x from $14k to $93k from September to October.
Sales & marketing used to be the bane of my existence as a technical founder.
Had to hammer it into myself until I was Stockholm syndromed.
2
5
298
Most people in the Bay don’t give a shit about privacy anymore the way people used to a few years ago.
Recording conversations is the de facto, and in some cases, actually expected.
But, the people who do care about it, care about it a lot. They get MAD about it.
Two questions:
1) how prevalent is this mindset in say, Tulsa, Oklahoma?
2) is the population of people who get mad going to increase or decrease in the next 2 years?
3
223
I was this close to building a property tax compliance B2B SaaS Al but I thought it was too “moonshot”
3
199
I’m so addicted to my stripe notifs now it’s not even funny I need to be less obsessed with revenue.
Dopamine hit from even a morsel more of MRR is crazy
feels like the era of hype is over.
signups, views, free users, github stars.. all meaningless vanity metrics
what matters now is revenue, margin, retention, and the sound of cash hitting your bank account
the tide is changing. and we're about to see who’s been swimming naked
3
190
That feeling when someone hits u w their Calendly link before u can.
Like bro I was just gonna send it to u why u gotta do my ego like that.
1
85
Genuinely curious, would you actually wear Mark Zuckerberg’s glasses with cameras & microphones on your face?
Feel like no one would but lots of people seem to be saying sentiment around Meta / privacy etc is getting better. Is it tho?
33%Would wear, meta is chill
67%F**k that
3 votes • Final results 1
1
93
Madhav Lavakare retweeted
AirCaps (@tryaircaps) is bringing AI assistance to real-world conversations.
Their software provides live captions, translations, and proactive AI insights on AR glasses. They’ve already processed 15,000 hours of conversations for hundreds of paying customers.