@MIT alum || Former Bridgewater Assoc. || Co-founder of @Modulate_ai
Joined May 2014
- Tweets244
- Following152
- Followers194
- Likes256
Michael Pappas retweeted
Building voice agents is half the problem.
Knowing what’s actually happening in those conversations is the other half.
In Atlanta for VON, demoing @modulate_ai at VoiceAI LIVE! today alongside @SoundHound and others.
2:40 PM, Room C. Come say hi. I’ll be listening to how you say it.
Michael Pappas retweeted
I used to teach spacecraft to see.
Now I teach software to listen.
@modulate_ai recently raised $25M, led by Future Ventures, to build the listening layer for AI.
Cosmology → JPL → voice. Here’s the path 🧵
Michael Pappas retweeted
Winnowing nuance from noise
Modulate runs over 100 voice-native models to discern emotion, tone, intent, and language, and detect synthetic and deepfake voices (ranked #1 on 🤗).
We just led their seed round; news today: techcrunch.com/2026/09/28/mo…
Product: modulate.ai/velma
We just raised $25M in new funding led by Future Ventures (@VenturesVc), with @HyperplaneVC and @lakestarvc - bringing Modulate’s total funding to $60 million.
Building the next phase of audio-native AI: more models, more developers, more of the stack that actually understands voice.
Where we stand today:
- Ranked #1 on Hugging Face for deepfake detection and transcription
- 600M+ hours of audio analyzed
- 10M+ hours analyzed every month
A foundational layer of the AI stack is getting more powerful. More soon.
Proud of this team.
$25M to keep building the understanding layer for voice AI - because the next generation of AI won't just talk to us, it'll actually have to understand us back.
Grateful to @VenturesVc @HyperplaneVC and @lakestarvc for backing that bet.
We just raised $25M in new funding led by Future Ventures (@VenturesVc), with @HyperplaneVC and @lakestarvc - bringing Modulate’s total funding to $60 million.
Building the next phase of audio-native AI: more models, more developers, more of the stack that actually understands voice.
Where we stand today:
- Ranked #1 on Hugging Face for deepfake detection and transcription
- 600M+ hours of audio analyzed
- 10M+ hours analyzed every month
A foundational layer of the AI stack is getting more powerful. More soon.
Michael Pappas retweeted
Started @modulate_ai with @mpappas74 because a transcript can't tell you if someone's about to hang up angry, or if a voice on the other end isn't human at all.
100+ specialized models later, we're #1 on @huggingface Face's deepfake speech benchmark and analyzing 10M+ hours of audio a month.
$25M to build a lot more of it.
We just raised $25M in new funding led by Future Ventures (@VenturesVc), with @HyperplaneVC and @lakestarvc - bringing Modulate’s total funding to $60 million.
Building the next phase of audio-native AI: more models, more developers, more of the stack that actually understands voice.
Where we stand today:
- Ranked #1 on Hugging Face for deepfake detection and transcription
- 600M+ hours of audio analyzed
- 10M+ hours analyzed every month
A foundational layer of the AI stack is getting more powerful. More soon.
Michael Pappas retweeted
You can't just dump recordings into a model and expect it to work.
Modulate CEO @whuffman on why voice AI lives and dies on data quality, not data quantity - and why context (language, region, setting) changes everything.
👇
hubs.ly/Q04xDKzs0
Michael Pappas retweeted
"Audio is not just what you said. It's how you said it." - Modulate Co-founder @mpappas74 on why voice deserves more nuance than we give it.
👇
Michael Pappas retweeted
Modulate CTO @whuffman during @ai4conferences: spoken words carry only 3–20% of the meaning in a voice conversation 👀
Tone, pace, and emotion are the rest - and none of it reaches a transcript-based LLM.
That's why new models alone won't fix a bad voice agent. It's an architecture gap.
Watch the full talk here: bit.ly/4yu26vw
Every voice AI team is having the same argument right now: cascade vs. speech-to-speech.
Most are arguing about the wrong thing.
The real question isn't architecture. It's what your STT layer throws away before your LLM ever sees it.
I wrote up why 🧵👇
Michael Pappas retweeted
🔉 We’re hiring a Backend Engineer at Modulate.
Help build and scale the Developer and Enterprise APIs powering our voice AI models. Ship code that puts ML audio models into production - not just papers.
Apply here: hubs.ly/Q04vH6HT0
Michael Pappas retweeted
6 reasons Modulate Transcribe is different from every other transcription API on the market:
1️⃣ It's #1 on the @huggingface Open ASR Leaderboard, out of 88 models evaluated on word error rate across real-world audio.
2️⃣ It's built on Modulate's Ensemble Listening Model (ELM) architecture: orchestrating multiple specialized transcription models together, instead of relying on one, for better accuracy, latency, and cost than any single model alone.
[🧵]
Michael Pappas retweeted
Still true: Modulate Transcribe is #1 out of 88 models on @huggingface Open ASR Leaderboard 🏅
Not our benchmark. Not our leaderboard. A public, third-party ranking anyone can check.
If you're evaluating transcription APIs, that's the kind of proof that doesn't need a press release to back it up.
See the ranking: bit.ly/4iagQuC
Michael Pappas retweeted
Modulate is partnering with @ScamAI_Official to help fraud teams see the whole attack, not just part of it.
ScamAI leads image, video, and document deepfake detection (98.2% accuracy).
We're bringing Velma, our synthetic voice detection model ranked 🏅1 on @Hugging Face (98.9% accuracy), into their platform.
One workflow. Every format scammers use.
Michael Pappas retweeted
Mandarin just got a new best-in-class transcription model. It's ours.
Modulate Transcribe posted the lowest Utterance Error Rate of any provider on Sierra's μ-Bench for Mandarin (zh-CN): 27.4%. Google Chirp-3 came in at 29.5%. ElevenLabs Scribe v2 at 37.1%. We beat both 🏅
And this wasn't the easy locale to win.
Sierra's own study named Mandarin the hardest in the benchmark - tone, segmentation, code-switching, the works. Most models take their biggest accuracy hit right here. We picked the hardest fight on the board and won it.
Quick context on why the number matters: UER is not WER.
It measures the share of utterances with at least one significant, meaning-changing error - the kind that actually breaks a voice agent or a downstream workflow. You can look fine on WER and still fail here. We didn't.
Then there's speed. Our P95 batch completion latency was 658ms. Chirp-3's was 1,196ms. That's 45% lower latency, on top of the lower error rate. Not a trade-off. Both.
Best accuracy on the leaderboard. Faster than Chirp-3. Hardest locale on the leaderboard.
Proud of the team for this! 🙌
Don't take my word for it. Benchmark us on your own audio: 1,000 free credits: bit.ly/4b6kS35
@huggingface just listed @modulate_ai as the best transcription model in the world. Best on which metric? Amazingly, nearly all of them.
We're the most accurate model. We're more than 10x *cheaper* than the runner-up models, and our models aren't just good on clear audio, but outperform on noisy, real-world content too.
Not to mention that we're the only provider that augments transcription with emotion, prosody, and behavioral understanding, to *truly* understand what's being said in a voice conversation.
huggingface.co/spaces/hf-aud…
How did we do it, while being literally 1000x smaller than many of our competitors? Aside from simply amazing talent, the secret is in the data. @modulate_ai got started on the hardest voice understanding problem out there - real-time understanding of video game voice chats.
That means being able to keep up with new jargon and phrasing. It means handling incredibly low-quality microphones as well as high-end headsets. It means understanding not just words but emotions, to tell the difference between friendly banter and genuine hate speech. And it means doing it all in seconds.
Others started with studio recordings, and produced technology that works in the studio. Modulate dove into real conversations, and built technology that leads the way in the real world.
Everyone's seen this scene from @netflix Queen Charlotte 👀
Here's what Velma by @modulate_ai sees.
Not just the words - but emotion, sentiment, conversational behaviors, speaker dynamics, and everything else encoded in the voice.
That's the difference between transcription and TRUE voice understanding.
Welcome to Velma in the Wild.
I'll be doing this every week, putting iconic clips through the world's most powerful Voice AI to show what conversations actually sound like beyond the transcript.
What should I analyze next?
#VelmaInTheWild
Michael Pappas retweeted
Ever misread someone just because of their tone? 😶
Now imagine that at scale - with AI.
In this clip, our CEO @mpappas74 gets into why voice isn’t just words. It’s pauses, emphasis, emotion - all the subtle signals people naturally interpret.
Miss that, and you’re misinterpreting the conversation itself.
That’s the layer we focus on at @modulate_ai - not just the spoke word, but the entirety of the conversation.
Michael Pappas retweeted
We’ve trained ourselves to question video.
But we still trust voice.
That’s what makes audio deepfakes so dangerous ⚠️
In this clip, our CEO @mpappas74 breaks down why voice is such a powerful attack surface - it’s how we verify identity, talk to banks, recognize people we know.
We don’t second-guess it. We assume it’s real 🤷♀️
And that’s exactly what bad actors exploit 🙅
At @modulate_ai, we’re building for that reality.