What you can do with a voice
Formanta verifies people by voice, detects voice clones, deepfakes and other synthetic voices, tells you who spoke when in a recording, and recognizes members of a household. Each use case below shows the business process step by step: what your system does, what Formanta does, and what you store.
Caller verification
Verify customers by voice when they call a contact center, talk to a voice assistant in your app or speak with a voice AI agent. Enroll each customer once, after consent. On every later call, check the voice in the first seconds of the conversation.
Why it matters
- Security questions and SMS codes add time to every call and are beaten with leaked personal data.
- A voice match alone can be fooled by a voice clone. The voice also has to be checked for liveness.
- A voice AI agent cannot approve a payment or change account data without knowing who is speaking, so these calls still go to a human.
What you get
- A voiceprint for each customer, stored in your own database next to the customer record.
- On each call, the probability that the voice is live, a clone or deepfake, or a replay.
- The probability that the caller is the enrolled customer, which you turn into verified, step-up or not verified.
Consent comes first. Voice used for authentication is biometric data under GDPR and similar laws. Obtain the customer's consent before enrollment, in the form your law requires. You collect and keep the consent record in your own systems.
- YouGet consentThe customer agrees to voice biometrics in the IVR or with an agent.
- YouRecord speechAt least 3 seconds of the customer's voice. 10 to 30 seconds is better.
- FormantaCheck it is liveReject replays, voice clones and bots before they become a voiceprint.
- FormantaBuild voiceprintA compact template of the customer's voice.
- YouStore itNext to the customer ID in your database.
- YouGet consentIn any channel: app, website, branch or call.
- YouSelect recordingsPast calls with the customer's voice, already in your archive.
- FormantaCheck each is liveRecordings that are not live speech are skipped.
- FormantaVoiceprint per callThen merge them. Speech from different calls makes a stronger voiceprint.
- YouStore itThe customer is enrolled without saying a single extra word.
A call recording holds both the customer and the agent. Use the customer's channel of a stereo recording, or run speaker diarization first to find the customer's speech.
- YouCapture speech3 seconds or more of the caller, plus the identity they claim.
- FormantaCheck it is liveStop here if the voice is a clone, a deepfake, a replay or a bot.
- FormantaCompareNew voiceprint against the stored one.
- YouDecideYour thresholds, your business rules.
With a voice AI agent: check before the agent acts
The agent already hears the caller, so the same audio can be verified while the conversation goes on. You choose when the check happens.
At the start of the call
Verify while the caller says hello and states the reason. The rest of the call runs as a verified session.
Before a sensitive action
Routine questions need no check. The agent verifies right before a payment, a password reset or a change of contact details.
Again during a long call
Send new audio as the call goes on, to catch a handover to another person after the first check passed.
Voice clone and deepfake detection
Find out who is on the line, a live person or a machine, without enrolling or verifying anyone. Formanta detects voice clones and other synthetic voices, voices converted in real time, AI voice agents and bots, recordings played back into the call, and suspicious injection paths.
Screen inbound calls
Score the first seconds of every call. Warn the agent, or route the call to a stricter process before anything sensitive happens.
Stop fake calls from "the boss"
A cloned voice of a manager asks finance to make an urgent payment. Check internal calls the same way you check customer calls.
Know when you talk to a bot
Voice AI agents and automated callers are flagged, so you can route them differently from people.
Warn people on their phone
With the on-device SDK, a phone or desk phone checks an incoming call locally and warns the user. The audio stays on the device.
- YouCapture speechA few seconds from the start of the call, or the whole recording.
- FormantaDetect spoofingA probability for each class below.
- YouAct on the resultContinue, warn the agent, ask for a step-up check, or end the call.
liveA live person is speaking.deepfakeA voice clone, a converted voice, a synthetic voice or an AI agent.replayRecorded speech played back into the call.On the call
Check the first seconds of a call, before the conversation reaches anything sensitive.
On recordings
Check full recordings as background jobs: insurance claims, recorded statements, voicemails, compliance reviews.
Before the voice match
A good clone can match the customer's voiceprint. Check liveness on the same audio first, then trust the match.
Speaker diarization
When a call or a meeting is recorded as one mono signal, Formanta finds how many people spoke and which parts each of them said. If you have voiceprints of the people who may be in the recording, it also tells you who they are.
Call analytics
Split agent and customer talk time, measure interruptions and silence, and run quality control on mono call recordings.
Readable transcripts
Label each line of a call or meeting transcript with its speaker, for note-takers, compliance records and search.
Enrollment from the archive
Find the customer's speech in a past call, so the agent's voice does not end up in the customer's voiceprint.
- YouSend the recordingA mono recording of the call or meeting. Long recordings run as a background job.
- FormantaDiarizeThe speakers and the time segments where each of them talks.
- FormantaVoiceprint per speakerBuilt from that speaker's segments.
- OptionalMatch known voicesCompare with your agents or enrolled customers. Otherwise they stay Speaker 1, Speaker 2.
Steps 1 and 2 answer how many people spoke and when. Steps 3 and 4 add names when you have voiceprints to compare with.
Personalization in voice assistants
Devices that several people use, such as a smart speaker at home, a car or a TV, recognize who is speaking and answer with that person's calendar, music, contacts or settings. Run it on the device, and the household's voices never leave the home.
Personal answers
"What's on my calendar" gives each person their own day. Playlists, contacts and reminders follow the voice.
Limits on commands
Allow purchases, unlocking or settings changes only for the people who should make them. A child or a guest gets a limited set.
The driver's profile
In a car, the system knows who is talking and loads that person's preferences, without a profile menu.
- YouCollect phrasesEach member says a few phrases, 3 seconds or more each, after consent.
- FormantaOne voiceprint eachPhrases are merged into one voiceprint per member and stored on the device or in your cloud.
- YouCapture the requestThe spoken command after the wake word.
- FormantaCompare with each memberThe best match above 0.5 is the speaker. Below that for everyone, it is a guest.
More from the same audio
Extra signals that work alongside any use case, from the same recording.
Signal quality
How much speech and how much noise a recording has, and the signal-to-noise ratio. Check a recording before you enroll from it.
Age and gender
An estimate of the speaker's age and gender, for analytics and for routing.
Emotions In development
Tell us how you would use it, and we will share progress.
Common questions
Do I need to enroll people to detect a voice deepfake?
No. Voice clone and deepfake detection works on any audio without a voiceprint. Enrollment is needed only when you also want to know who is speaking.
How much speech does Formanta need?
One second of net speech is the minimum, and three seconds or more is recommended. Accuracy keeps improving up to about 30 seconds of speech, especially when the speech comes from different calls, which you can merge into one voiceprint.
Does it work on phone calls?
Yes. Telephone audio, which is usually recorded at 8 kHz, is supported.
Do people have to say a passphrase?
No. Verification is text-independent. People can say anything, in any language or accent, so the check runs during the normal conversation. Instead of relying on a fixed phrase to resist recordings, every request can include a liveness check.
Can a voice clone pass voice verification?
A good clone can sound close enough to match the customer's voiceprint. That is why Formanta checks the same audio for voice clones, deepfakes and replays first, and you only trust the voice match when the voice is live.
What if a genuine customer has a cold or calls from a noisy street?
The score can drop. Use an unclear zone around the threshold and send those calls to a step-up check, such as a callback or an app confirmation, instead of rejecting the customer. Adding speech from later calls to the stored voiceprint keeps it current.
Can Formanta find a caller among many people?
Today you can compare a voice with a small group of voiceprints, such as the members of a household or a team of agents. Talk to us if you need to search a large database.
Can a voiceprint be turned back into a voice?
No. A voiceprint is a set of numbers that describes a voice, not a recording, and it cannot be played back.
Where is voice data stored?
Voiceprints are returned to you and stored in your own database. In the cloud, each organization sees only its own data, and job results are deleted 24 hours after the job finishes. On-premise and on-device, audio does not leave your network or the device.
Do I need consent from the people I enroll?
Yes. A voice used to identify a person is biometric data under GDPR and similar laws, so you need the person's consent in the form your law requires. You collect and keep consent in your own systems.
How do I get access?
Use the form below to ask for early access to the cloud API and the MCP server, or to join the on-device SDK beta. On-premise deployment is part of the enterprise plan.
Is the API final?
The Voice Server API v1.0 is a draft. Endpoints and fields may change before the public launch of the cloud API and the MCP server.
Try it on your own audio
Tell us which use case you need and where your audio is. We will show the results on recordings from your environment.