There's a lot of anxiety in our industry right now, and it's not unfounded. Generative AI has made synthetic speech cheap, fast, and good enough to fool a casual listener — and in some documented cases, good enough to fool a bank. Every few months brings a new headline: a journalist cloning his own voice to break into his own account, a finance employee wiring millions after a deepfake video call, a fraud ring using off-the-shelf tools that didn't exist three years ago. It's reasonable to ask whether voice biometrics can survive this.

We think it can. But not in its original form.

The Problem Was Never Voice. It Was the Single Check.

Traditional voice authentication asks one question, once: is this the right person, right now? That model made sense when the hardest part of an attack was getting a convincing sample of someone's voice in the first place. It doesn't make as much sense today, when a few seconds of audio scraped from a podcast, a voicemail, or a customer service call can be enough to generate a passable clone.

Here's the part that gets missed in most of the “voice biometrics is dead” takes: a single check at login was always a narrow target, and narrow targets are easy to plan around. If a fraudster only has to fool the system once, at a predictable moment, they can prepare for that moment specifically. The vulnerability was never the biology of voice. It was giving an attacker exactly one door, at a time they could see coming.

Continuous Authentication Closes That Door — But Not by Itself

The fix isn't to abandon voice as an authentication factor. It's to stop treating authentication as a single event. Continuous authentication verifies the caller throughout the interaction, not just at the start — an IVR check-in, a contact center call, a video-conferenced approval. If a spoofed or hijacked voice takes over mid-call, the system doesn't wait for the next login to notice.

This alone changes the economics of an attack. Fooling a verification system for two seconds at login is a very different problem than sustaining a convincing forgery for a five-minute conversation, under the exact acoustic conditions the model expects, without a single break in the pattern.

But continuous authentication on its own answers the wrong question. Matching a voice against a stored voiceprint tells you the pattern is consistent — it doesn't tell you the audio is genuinely human in the first place. A sufficiently good synthetic voice, generated consistently, could in principle produce a stable match throughout an entire call. Voice matching and deepfake detection aren't redundant systems solving the same problem twice; they're answering two different questions that both need to be “yes.” Is this voice consistent with the person on file? and, separately, is this actually a human voice at all, and not a synthetic one? You need both running continuously, in parallel, for the whole call — not one covering for a gap in the other.

Voice Matching and Deepfake Detection Aren't the Only Signals Worth Watching

Once you're already analyzing a call continuously, the marginal cost of layering in additional fraud signals is low — and the additional signal is worth having. A few examples, alongside voiceprint matching and anti-spoofing:

• Audio quality and channel characteristics. Synthetic audio, recordings, and re-transmitted calls often carry subtle artifacts that differ from a live, direct conversation. Flagging those anomalies doesn't prove fraud on its own, but it's a useful input alongside everything else.
• Acoustic classifiers trained to estimate characteristics like age range or vocal register from a sample. It's worth being precise about what these actually measure: acoustic and vocal-tract characteristics statistically correlated with certain populations — not a verified, self-reported identity attribute. Used well, this is one more consistency check: does this call's acoustic profile look like what's on file for this account, or has something changed that's worth a second look?
• Behavioral and emotional tone. Sudden shifts — unusual stress patterns, a change in speaking pace or register mid-call — can be one more input into a risk score, the same way a fraud team already weighs unusual transaction timing or an unfamiliar device.

None of these signals is a fraud verdict by itself, and none should be treated as one. The point isn't to build a single classifier that claims to know who someone is from their voice alone. It's defense in depth: several independent, imperfect signals, each catching something the others might miss, combined into a system that's much harder to fool than any one check would be on its own.

What This Looks Like in Production

This isn't theoretical for us. Our Twilio Media Streams connector now runs continuous authentication for the full duration of a call, not just at connection — combined with real-time deepfake detection running the entire time, not sampled at intervals. The two run together because they have to: continuous voice matching without continuous anti-spoofing just means an attacker has more time to get the forgery right, not less incentive to try.

The layered-signal approach matters more as the baseline threat gets better. Anti-spoofing that catches today's generative models needs to keep improving as those models do — which is exactly why we don't treat deepfake detection as a solved, static feature. It's a moving target, and the systems watching for it need to move with it.

Voice Still Makes Sense as a “Who You Are” Factor — Used the Right Way

Despite everything above, we haven't lost confidence in voice as an authentication factor. If anything, the case for it is stronger than most other biometrics, for one specific reason: voice is the one factor that can be verified passively and continuously, throughout a natural interaction, without asking someone to do anything differently. You can't continuously re-scan a fingerprint during a phone call. You can continuously listen.

The synthetic speech problem is real, and it deserves the scrutiny it's getting. But the answer isn't to retreat from voice biometrics — it's to stop asking a single yes-or-no question and start running a system that keeps asking, and keeps watching for the signs that the answer has changed.