Skip to content
Rush Commerce
AI & Automation3 min read

Modulate's $25M: your voice agent needs a listener

Modulate raised $25M for audio-native models that detect deepfake callers and score how voice agents are actually performing. The listening layer is the product.

Everyone shipping a voice agent is measuring the wrong side of the call. You watch containment rate and call length. You do not watch the raw audio for a caller who is quietly furious, or for a voice that was synthesized ten seconds ago. Modulate raised $25 million to sell that second half. TechCrunch reported the round on September 28, led by Future Ventures with Hyperplane and Lakestar participating, bringing total funding to roughly $60 million for the Boston company.

What actually happened

Modulate's models work on raw audio rather than a transcript. Emotion, tone, language, and synthetic-voice artifacts all register as signals, and a second layer turns those signals into judgments: intent, rule violations, scam patterns, deepfake attempts.

The architecture is the interesting part. Per SiliconANGLE's coverage, an Ensemble Listening Model picks from more than 100 small specialized audio models per job and blends the results, which Modulate says is up to 1,000 times more efficient than handing the same audio to one large model. The company reports more than 10 million hours of audio a month running through it, two first-place finishes on Hugging Face leaderboards this year, and 98.9% accuracy for its deepfake model on public benchmark data. Those last figures are the vendor's own; treat them as claims until you test on your calls.

Modulate started in game voice-chat moderation. Healthcare institutions now use it to screen callers impersonating staff with cloned voices, and a newer cohort of customers points it at their own voice agents to check performance. CEO Carter Huffman made the point that ordinary sentiment scoring misses: people stay polite to a bot and are dissatisfied anyway.

Why voice agent monitoring matters for your business

If you put an AI receptionist on your main line this year, you added two exposures and probably instrumented neither.

The first is inbound fraud. Voice cloning is cheap and your front desk — human or agent — is the softest identity check you own. Any workflow where a caller can move an appointment, change a shipping address, reset access, or authorize a refund on voice alone is now a vector. The fix is not a better model; it is a policy. Decide which actions a phone call can trigger without a second factor, and make everything above that line require a callback to a number on file or a link to an authenticated session.

The second is silent failure. Containment rate tells you the agent finished the call. It does not tell you the caller gave up and never came back. Sample your own recordings weekly — twenty calls, listened to by a person — and score them against the outcome you actually wanted: booked, resolved, escalated cleanly. Politeness is not satisfaction, and a transcript hides the pause where someone realized they were talking to a machine about something that mattered.

You do not need Modulate's stack to act on this. You need a written policy on what voice can authorize, and a weekly habit of listening to calls your dashboard called successful.

Key takeaways

  • Modulate raised $25M led by Future Ventures on September 28; roughly $60M raised in total
  • The models read raw audio — emotion, tone, synthetic-voice artifacts — rather than a transcript
  • Modulate reports 100+ specialized models in an ensemble, 10M+ hours of audio monthly, and 98.9% deepfake accuracy on public benchmarks; these are vendor figures
  • Voice cloning makes your phone line an identity-verification problem, not just a support channel
  • Define which actions a call can authorize without a second factor; require callback or an authenticated link above that line
  • Containment rate hides abandonment — sample and listen to twenty "successful" calls a week

A voice agent you cannot audit is a stranger answering your phone. We build phone and chat automation with explicit limits on what a caller can authorize, callback verification on anything that moves money or access, and recordings scored against real outcomes instead of containment rate. See how we build voice automation you can audit, or tell us what your front desk currently says yes to.

Sources: TechCrunch, SiliconANGLE.

  • #voice-ai
  • #ai-agents
  • #fraud-detection
  • #funding
  • #monitoring
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.