Modulate Raises $25 Million to Catch Deepfake Voices Mid-Call, at $0.25 an Hour of Audio

Modulate has raised $25 million to expand a platform that listens to voice calls and works out what is actually happening in them. The Boston company announced the round on Sept. 28; Future Ventures led, with Hyperplane and Lakestar participating. PitchBook data cited by TechCrunch put Modulate's prior funding at $41 million and its last valuation at about $170 million. Total funding is now roughly $60 million, and the company says it employs between 40 and 45 people, with plans to add about ten more.

The product is Velma, an audio-native platform built on what Modulate calls an Ensemble Listening Model architecture. Rather than finishing with one large foundation model, ELM orchestrates more than 100 narrow ones. Some extract signals such as emotion, tone, emphasis and whether a voice is synthetic; others judge higher-level behavior such as whether a caller is attempting a scam or a voice agent is violating a compliance rule. Signals can be used separately or stacked to recognize events.

That design choice carries the company's commercial argument. Small models do not need specialized hardware, and chief executive Carter Huffman argues they make it cheaper to add a capability as new attack techniques appear: a new model joins the ensemble and an orchestrator calls it when relevant. Modulate says the approach has demonstrated up to 1,000x greater efficiency than a single large-model pipeline for the same audio analysis. That is a vendor claim and has not been independently verified.

The scale figures are the company's own as well. Modulate says it processes more than 10 million hours of audio a month and has handled over 600 million hours cumulatively, and that its transcription and deepfake-detection models both sit at the top of Hugging Face's public leaderboards. On benchmark data the company cites, deepfake speech detection reaches 98.9% accuracy with a 1.1% equal error rate, and Velma is claimed to be twice as accurate as traditional LLMs applied to audio, with seven times fewer false positives. Those numbers come from Modulate rather than from an independent evaluation.

Pricing is public and unusually granular: $0.25 per hour of audio for the deepfake detection API and $0.03 per hour for batch transcription. For enterprises weighing deployment, that is the difference between a control applied to every call and one applied to a sample.

Modulate was founded in 2017 by Carter Huffman and Mike Pappas, who met as MIT physics undergraduates, and started out selling voice modulation for games, essentially voice skins that let players sound like someone else. The pivot to detecting manipulated voice came later, and the company now sits beside a customer's existing voice stack, analyzing calls rather than producing them. It counts healthcare institutions, call centers and social platforms among its users, and says the same models are used to flag child grooming in voice conversations and to protect agents in high-risk scenarios from being identified by voice.

"Voice is becoming a primary interface for AI, and that creates a whole new set of problems that can't be solved from a transcript," Huffman said. The point is technical as much as rhetorical: transcription discards tone, emphasis and the small inconsistencies that separate a recording of a person from a model trained to imitate one.

The funding is a bet that voice fraud matures on a curve similar to email phishing, from nuisance to line item, and that a detection layer ends up sitting below the CRM and call-center software enterprises already run. Modulate says it is also building out on-premises and on-device deployment, which matters for healthcare and financial customers unwilling to stream live calls to a third party. What the market still lacks is an independent measure of how these detectors perform when an attacker is specifically optimizing against them. The accuracy numbers in circulation today come from benchmarks, not contested calls.