A critical assessment of the $25M Future Ventures-led raise for “audio-native” voice AI — best known for ToxMod, the moderation tool that listens to voice chat in Call of Duty, now the subject of a biometric-privacy investigation, with self-graded performance numbers and a core emotion-inference product sitting squarely in the EU AI Act's crosshairs.
Unclear. TechCrunch declined to label the round; secondary outlets call it a Series B. No new valuation was released, and total-raised figures don't reconcile — ~$41M, “$60M,” and ~$66M appear across three sources. The clean “up-round” narrative is partly manufactured by the press.
Emotion and intent inference from voice — not transcripts. That is the differentiator and the liability. The EU AI Act restricts emotion-recognition systems, and biometric rules are tightening globally, yet Modulate frames regulation only as a demand tailwind, never as a constraint on its own tech.
The only independently verifiable wins are narrow Hugging Face leaderboard ranks on academic tasks. The headline numbers — “2x better than LLMs,” “98% accuracy,” “1,000x more efficient” — are all Modulate grading its own homework, not live false-positive rates.
Key Finding: Modulate has something rare — proprietary, self-built audio models (not a wrapper on OpenAI or Anthropic) and a real, verifiable marquee customer in Activision's Call of Duty. But its expansion pitch rests on emotion and intent inference, the exact capability regulators are moving to restrict; its marquee deployment is under a BIPA biometric-privacy investigation; and nearly every performance claim is self-benchmarked. Independent journalists describe the product with a blunter verb than the company does: it listens.
Modulate's thesis: you can't solve voice's problems from a transcript. Its models work on the raw audio — extracting signals a transcript throws away. That capability is the moat and the regulatory exposure at once.
Infers emotional state from vocal signals — the exact capability the EU AI Act moves to restrict.
Classifies what a speaker is trying to do — harassment, grooming, fraud — from how they say it.
Flags deepfaked or AI-generated speech; ranks #1 on the Hugging Face Speech Deepfake Arena.
ToxMod surfaces bad actors in real time — live across Call of Duty since November 2023.
The company says “analyze.” Independent journalists say “eavesdrop” and “listen.” Both describe the same thing: software processing tens of millions of hours of players' live speech every month. Modulate's defense — “we generate no voiceprints, do no speaker ID” — is a narrow, self-defined technical claim no court or regulator has yet tested.
A consumer-rights investigation alleges that Activision, via ToxMod, may be capturing and processing Call of Duty players' voice data without the notice, consent, and retention disclosures that Illinois' Biometric Information Privacy Act (BIPA) requires. Status matters: this is an investigation soliciting plaintiffs for mass arbitration — not a filed lawsuit. But BIPA carries statutory damages per violation, and the exposure follows the product into every deployment. Modulate's rebuttal is a technical definition; whether it satisfies the statute is untested.
Steve Jurvetson's fund led the round, framing audio intelligence as “a foundational layer of the AI stack.”
A genuine differentiator: Modulate builds its own models on AWS — not reselling OpenAI or Anthropic. Its only hard dependency is cloud.
Real, third-party-verifiable leaderboard wins (ASR, Deepfake Arena) — but they measure narrow academic tasks, not messy live audio.
The marquee, independently confirmed customer — and the source of concentration risk if Activision ever in-houses moderation.
New flagship platform claimed “2x as accurate as general-purpose LLMs” with “7x fewer false alarms” — Modulate's own comparison.
The expansion engine — and the regulatory blast radius. The EU AI Act restricts emotion-recognition systems in multiple contexts.
Modulate makes big, specific performance claims. The independently checkable ones are narrow; the impressive ones are self-reported.
Third-party signals are real but narrow: #1 on the Hugging Face Open ASR Leaderboard and Speech Deepfake Arena (1.1% equal error rate). The Activision/Call of Duty deployment is confirmed by Activision's own support pages and multiple outlets. Vendor independence checks out — the privacy page confirms AWS, names no LLM vendor.
“10M+ hours of audio monthly,” “600M hours analyzed,” “98% accuracy in each category of harm,” “2x LLMs / 7x fewer false alarms,” “1,000x more efficient” — every one is company-supplied and unaudited. No ARR or revenue is disclosed by anyone. Leaderboard rank is not live moderation accuracy, which is exactly where gamers historically complain.
The funding math itself doesn't reconcile. TechCrunch puts total raised at ~$41M; the company's release says “$60M”; PitchBook figures imply ~$66M. Three sources, three numbers — and the new-round valuation was withheld entirely. A flat or soft round dressed as an up-round is one reading the public record cannot rule out.
Seven structural risks the $25M raise does not resolve.
The core “vocal emotion / intent” capability collides with EU AI Act restrictions on emotion-recognition systems and tightening biometric rules. A “we make no voiceprints” defense may not cover emotion inference at all.
An active investigation into ToxMod-in-Call-of-Duty (mass-arbitration route) means biometric-privacy exposure follows the product into every U.S. state deployment — even before any suit is filed.
Marquee revenue appears tied to Activision and a handful of large platforms — undisclosed and undiversified. Any one in-housing moderation could gut revenue.
Headline accuracy and efficiency claims are all internal; the independent benchmarks don't measure the live false-positive rates that drive gamer backlash.
ActiveFence (which acquired Spectrum Labs; valued >$500M) and expanding ASR/voice-agent players out-resource Modulate as it pushes beyond gaming into security and CX.
Chasing moderation, deepfake detection, transcription, fraud, child-safety and voice agents at ~45 people risks spreading a small team thin against focused specialists.
Irreconcilable total-funding figures, an undisclosed new valuation, and an unexplained mid-round CEO handoff make capital efficiency and runway hard to read from the outside.
Modulate is one of the few voice-AI companies that actually owns its models — a real edge as the field fills with LLM wrappers. But the same capability that makes it valuable, inferring emotion and intent from your voice, is the one regulators are moving to restrict, and its marquee deployment is already drawing biometric-privacy scrutiny. The key diligence questions are not technical: they are whether the regulatory ground shifts under emotion recognition, and how much of the revenue rides on a single gaming customer that could one day build this in-house.
Based entirely on publicly available information, including the TechCrunch announcement of September 28, 2026. Company-supplied metrics are treated as claims; the BIPA matter is an investigation, not a filed lawsuit.
We build and maintain websites, help businesses of every size put AI to work, and run growth marketing that’s held to the same standard as our research.
