←
Carter Huffman149 posts
Carter Huffman avatar

Carter Huffman

@whuffman

@MIT alum | Formerly @NASA @NASAJPL | Current CEO & Co-founder of @Modulate_ai

◷ Joined August 2015▧ 65 Photos and videos
Carter Huffman avatar
Carter Huffman@whuffman·Nov 10

1.6 seconds to tell a synthetic voice from a human, median. Down from 2.25. The harder problem is telling a helpful AI caller from a malicious one. Thanks @twilio Ventures and @englishpaulm for backing us.

@twilio Ventures + Kayak co-founder Paul English (@englishpaulm) are now Modulate investors. First update since: median synthetic voice detection 2.25s → 1.6s. Accuracy >99.1% on real-world calls. Read the press release below 🧵

Post attachment
Carter Huffman avatar
Carter Huffman@whuffman·Nov 10

ICYMI: Had a great conversation with @donlterry yesterday about the growing challenge of voice fraud and deepfakes! AI-generated voices are becoming so convincing and knowing who’s actually on the other end of the line is also becoming a much more interesting challenge. How do we distinguish real customers from impostors without adding friction for everyone else? We explored that and more! Catch the replay⬇️ @modulate_ai linkedin.com/video/event/urn…

Post attachment
Modulate avatar
⟳ Carter Huffman retweeted
Modulate@modulate_ai·Nov 10

Synthetic ≠ suspicious An AI booking your dinner reservation and a deepfake impersonating an employee both sound synthetic. Velma layers in intent, manipulation, tone, and behavior so teams know which is which.

Post attachment
Carter Huffman avatar
Carter Huffman@whuffman·Nov 10

Undergrad me: working out how the early universe evolved and what traces it left that we can see today. Current me: working out whether a caller is frustrated. Both are the same problem: infer what’s hidden from the signal that reaches you. One of them is slightly easier. I won’t say which.

Carter Huffman avatar
Carter Huffman@whuffman·Nov 9

Velma in the Wild, Ep. 4: Euphoria, S1. This one’s all about performance. Watch Velma track it straight from the audio: frustration → anger No transcript. Just voices. Words give you the script. Voice gives you the performance. What should we analyze next?

1:31
Carter Huffman avatar
Carter Huffman@whuffman·Nov 9

Every company building a voice product hits the same wall: The words are the easy part. Understanding is not. Emotion. Intent. Deepfakes. Who’s speaking. Whether the caller is getting angrier. You can become an audio AI research lab… …or plug into one. 1,000 free credits, no card. Go build the actual product 👇

Post attachment
Carter Huffman avatar
Carter Huffman@whuffman·Nov 9

AI-generated voices are becoming so convincing and knowing who's actually on the other end of the line is also becoming a much more interesting challenge. I believe it's not just about catching fraud (although that is very important!)...but I think it's just as important to do so without making life harder for real customers! I'll be chatting with our CRO, @donlterry, on @LinkedIn Live today about what we are seeing in voice fraud, why transcripts only tell a fraction of the story and how we can approach detection in real time. We'll also preview a few ideas I'm bringing to VapiCon later this week -- excited!! Hope you'll join us at 3PM ET. @modulate_ai linkedin.com/events/defendin……/theater

Post attachment
Carter Huffman avatar
Carter Huffman@whuffman·Nov 6

Voice is some of the most personal data there is. We don’t train on customer conversations. Customers control how their audio is used. That should be the baseline for every voice AI company, not a feature.

Carter Huffman avatar
Carter Huffman@whuffman·Nov 5

Our English Fast Transcription runs batch jobs at over 200x real-time. That’s an hour of audio in under 18 seconds. Your coffee takes longer. And 1,000 free credits covers 400 hours of transcription. No card required. 👇

Carter Huffman avatar
Carter Huffman@whuffman·Nov 5

A theme keeps emerging in conversations about #AI: as technology accelerates, human connection becomes more and more valuable. Tonight’s GTM discussion at @HyperplaneVC (as part of @TechCrunch Founder Summit 2026) reinforced that. Better data, smarter tools, faster execution. Yet trust, the power of connection & relationships, and deeply understanding your customers remain fundamental in the AI economy. The opportunity is in getting BOTH right.

Post attachment
Carter Huffman avatar
Carter Huffman@whuffman·Nov 4

COVID touched almost everyone, and many people lost someone they loved. One question I keep coming back to: could the sound of someone’s voice have helped us understand the spread earlier? Some research suggests voice may carry signs of illness. It’s early, far from proven, and the caveats matter: accuracy, consent, privacy. But if voice technology can one day help people get care sooner, it’s worth building carefully and with people’s trust.

0:29
Modulate avatar
⟳ Carter Huffman retweeted
Modulate@modulate_ai·Nov 2

#1 on @huggingface ’s Speech DF Arena 🏅 Modulate Velma ranks first on the independent deepfake detection leaderboard, with a 1.1% average equal error rate. The next-best model sits at 2.1%. Equal error rate (EER) measures how well a model separates genuine human speech from AI-generated audio. Lower is better. Why it matters for fraud teams: a cloned voice on a live call doesn’t wait for post-call review. Velma flags synthetic voices in real time, alongside other voice-native fraud signals: → Voice mismatch against known account holders → Scripted or rehearsed speech patterns → Emotional incongruence, like urgency without stress Every flag ties back to the original audio, so decisions hold up in audits and disputes. And it does this with 316M parameters. The two models ranked behind it on the leaderboard run at 1B and 3B. See how Velma stops voice fraud: hubs.ly/Q04BwQTK0

Post attachment