Modulate Cuts AI Voice Detection to 1.6 Seconds, Adds Twilio Backing

Modulate says its latest models can identify an AI-generated voice in a median of 1.6 seconds, down from 2.25 seconds, while keeping detection accuracy above 99.1% on its real-world evaluation of AI and human calls. The Boston-area company announced the numbers alongside new investments from Twilio Ventures and Kayak co-founder Paul English. It did not disclose the size of the additional investments.

The fresh backing follows Modulate's $25 million financing announced in September, led by Future Ventures with participation from Hyperplane and Lakestar. Alongside the money, the company is adding more than a dozen audio-native signals to its Velma platform. The goal: help enterprises understand both who — or what — is calling, and how an AI-led conversation is unfolding.

That distinction is becoming central to voice automation. A synthetic caller might be a helpful assistant booking an appointment, or a cloned voice attempting impersonation. Detecting generated speech supplies one piece of evidence. Deciding what to do with the call requires a broader view of its behavior and context.

Why does detection speed matter mid-call?

The reduction in median detection time matters because of when it happens. At 1.6 seconds, businesses can identify synthetic speech early enough to route a call, flag it for review, or apply extra verification while the conversation continues.

Modulate's system examines the underlying audio directly. It does not depend on watermarks inserted by individual voice-generation providers — including for newer speech-to-speech and full-duplex models.

The change from 2.25 to 1.6 seconds represents about a 29% reduction in the company's reported median. It is a detection-time measure, not a guarantee that every call will be classified within that interval, or that a full security workflow will finish that quickly. The announcement does not provide the evaluation sample size, or a breakdown of false positives and false negatives behind the greater-than-99.1% accuracy figure.

"Call operators – human or AI agent – must be able to recognize what's happening in those conversations and respond appropriately while the interaction is still ongoing," said Carter Huffman, CEO and co-founder of Modulate.

For contact centers, the operational benefit is intervention before a conversation ends. A warning delivered after an account change or sensitive disclosure has already occurred serves a different purpose from a signal available during the interaction itself.

How does Velma combine its models?

Velma uses what Modulate calls an Ensemble Listening Model, or ELM. The architecture coordinates specialized models rather than relying on a single general-purpose model to interpret a transcript. The company says the system brings together more than 100 models covering signals such as:

• emotion
• tone
• intent
• cadence
• synthetic speech
• conversational behavior

An orchestration layer combines lower-level signals into higher-level assessments while preserving links to the underlying evidence. Time-stamped events can indicate where a conversation changed. That gives an operations team something more specific to review than an unexplained call-level risk score.

The design choice reflects a basic limit of text. A transcript captures words, but it does not preserve every pause, interruption, change in emphasis, or delivery pattern. Modulate adds acoustic information to the analysis, so developers can examine how a conversation proceeds alongside what participants say.

The company's deepfake-detection documentation illustrates the distinction at the API level. It describes acoustic classification of audio segments, with time boundaries, verdicts, and confidence scores. It supports both recorded files and live streams. A detection confidence score is evidence for a classification — not proof of a caller's identity or intentions.

Is synthetic speech the same as fraud?

Modulate explicitly separates these questions in its voice-fraud screening guide. Deepfake detection evaluates whether a voice is synthetic. Conversation analysis evaluates behavior. A legitimate AI assistant can use generated speech. A human caller can attempt social engineering without any voice synthesis at all.

The distinction also matters for people using assistive speech technology. A synthetic-voice flag alone is an insufficient reason to treat a caller as malicious. In the new release, Modulate says Velma can combine that flag with signals associated with intent, manipulation, tone, and conversational behavior to support more informed handling.

The practical goal is a graduated response:

• ordinary routing for permitted automation
• additional verification where a transaction warrants it
• escalation when the conversation presents suspicious behavior

The detection system supplies evidence. The organization's policies determine the response.

Can Velma supervise your own AI agents?

The new signals also target a less dramatic but commercially important problem. An AI agent can sound fluent while delivering a poor service experience. Modulate says the additions help evaluate response timing, audio quality, and conversational misunderstandings that transcript-only analytics can miss.

On its AI guardrails page, the company describes Velma as monitoring participants rather than acting as the conversational agent itself. Customers can configure monitoring around operating procedures and expected behavior, then send signals into existing systems through APIs and webhooks.

That separation matters. The system doing the talking is distinct from the system assessing the interaction. It gives teams a way to investigate where an agent is causing frustration, going outside policy, or failing to respond appropriately — and use those findings to improve deployment.

What does the Twilio connection signal?

Twilio Ventures focuses on AI, developer tools, and customer-engagement businesses, which puts audio intelligence squarely within its investment remit. Modulate already has a technical connection to Twilio. A May 2025 announcement described compatibility for its VoiceVault and ToxMod products, including Twilio Media Streams-compatible APIs for bringing live call audio into its systems.

The new investment builds on that history, though the announcement does not establish a new exclusive distribution agreement. Modulate says the capital will support research, product development, developer adoption, partnerships, and broader deployment options.

September's funding announcement placed the same priorities at the center of its expansion beyond earlier voice-safety applications. With this update, the company is bringing detection and supervision closer together — helping enterprises recognize generated speech, examine the surrounding behavior, and assess whether their own voice agents are doing the job intended.

The stakes are straightforward. As AI voice agents move from demos into production call flows, the enterprises deploying them will need to answer two questions at once: is this caller real, and is my agent behaving? Modulate is betting that neither question can be answered from a transcript alone.