Trust layer for voice AI
Who is speaking, and is the voice live and human? Formanta answers both from natural speech, in the cloud, on your servers or on the device.
Voice is everywhere. Voice identity is not.
Voice is the most natural way to talk to AI agents, bots and devices. Most voice interfaces still cannot tell who is speaking. Where they can, a conventional voiceprint may accept a cloned or converted voice.
No personal answers
The interface cannot personalise the dialogue, find the right context or apply the user's permissions.
Easy to spoof
Voice cloning makes a plain voiceprint one of the easiest biometric identities to fake.
Back to PINs and codes
People are pushed to a PIN, a one-time code, security questions or another channel.
A trusted voice interface has to know who is speaking and that the voice is live and authentic. Formanta keeps both checks inside the natural conversation.
Voice became an interface, and lost its implicit trust
Conversation becomes the interface
Voice AI, bots and smart devices move service, commerce and assistance into natural conversation.
Synthetic voice is ordinary traffic
Platforms have to tell an authorised AI agent from an impersonator, and a live person from a replay or an injected stream.
Cloud voice ID services were retired
Major cloud providers shut down their speaker recognition services in 2025 and 2026. Their customers need a modern replacement.
See the migration path →Regulators already treat synthetic voice as a fraud channel: a FinCEN alert on deepfake fraud schemes (November 2024), an FCC ruling on AI-generated voices in calls (February 2024), and transparency duties under Article 50 of the EU AI Act, which apply from August 2026.
One trust layer for every voice interaction
Formanta checks the same audio for identity and authenticity and returns one result your rules can act on.
Speaker verification
Text-independent and language-agnostic verification from natural speech. No passphrase.
Replay and deepfake defence
Detects replayed recordings, synthetic speech, voice conversion and suspicious injection paths.
Policy-ready output
One risk result for login, payments, personalisation and the permissions of an AI agent.
Judge it on your own audio
A benchmark score depends on the data set, the channel and the threshold, and a single number is easy to misread. We share detailed results with partners and customers, and run Formanta on recordings from your environment so you can see the scores before you decide.
Speaker verification
Measured on public speaker verification benchmarks. Detailed results are available on request.
Anti-spoofing
Evaluated separately from verification: a strong voiceprint score says nothing about how well a system catches replays, synthetic speech or converted voices.
A pilot on your recordings
Lab audio is not your phone line. A pilot shows how Formanta performs on your channel, languages and customers.
Ask for a pilot →Four jobs one voice engine does
Each use case runs on the same audio and the same models. Start with the one that matches your problem and add the others when you need them.
Caller verification
Confirm that a caller is who they claim to be before a login, a payment or an action by an AI agent.
Voice clone and deepfake detection
Tell a live person from a voice clone, converted voice, AI agent or played-back recording.
Speaker diarization
From a single mono recording, find how many people spoke, which parts each said, and who they are if you know their voices.
Personalisation
A smart speaker, a car or another shared device recognises who is speaking and answers with that person's settings.
Voice analytics from the same audio
Extra signals you can request next to any use case.
Run it where the conversation runs
The same models work in the cloud, on your servers and on the device.
Cloud API and MCP server
A REST API for your backend, and an MCP server that gives AI agents a native tool for voice checks.
On-premise
The same API in your own infrastructure. Audio never leaves your network.
On-device SDK
Phones, PCs, desk phones, smart speakers and cars check the speaker and the voice locally. Audio and voiceprints can stay on the device.
Enroll, check, compare, decide
A voiceprint is a numeric profile of a voice, not a recording. Formanta builds it and hands it back to you, and you store it next to the user record.
Enroll
After the person consents, 10-15 seconds of net speech become a voiceprint. Speech from several calls can be merged.
Check
New audio is checked first: live person, replay, synthetic speech or a converted voice.
Compare
A few seconds of the conversation are compared with the stored voiceprint.
Decide
Allow, ask for a step-up check, or refuse, by your own thresholds.
The name Formanta comes from formants: the resonant frequencies of the vocal tract that make each voice distinct.
Read the voice biometrics guide: attacks, metrics, when to check →An API call, or a tool for your agent
Call the REST API from any backend: send WAV audio, get JSON back. Or connect an AI agent through the MCP server, so it can check the voice before it pays, books or changes data.
def verify(user_id, wav):
if not is_live(wav):
return "rejected" # clone, deepfake, replay or bot
p = match(db.load_voiceprint(user_id),
voiceprint(wav))
if p >= 0.6: return "verified"
if p >= 0.4: return "step_up" # extra check or agent
return "not_verified"What faster authentication is worth
Replace security questions with a check that runs while the caller talks. Change the numbers to match your contact centre.
Estimate covers agent time only. Fraud losses avoided come on top.
You decide where voice data goes
Voice used to identify a person is biometric data under GDPR and UK GDPR. Formanta is built so that voiceprints stay with you, and so that audio can stay in your network or on the device when it has to.
Consent and retention checklist →Voiceprints stay with you
The service does not store voiceprints. It returns them to you, and you keep them in your own database.
Short retention in the cloud
Each organisation sees only its own data. Results of background jobs are deleted 24 hours after the job finishes.
Nothing leaves your network
On-premise and on-device, processing happens on your servers or on the device. Nothing is sent to Formanta.
A profile, not a recording
A voiceprint is a set of numbers and cannot be played back.
Enrol once, verify across agents and devices, while the voiceprint stays under the user's control.
We are launching soon
Formanta is built by Qualia Inc. We are preparing the cloud API, the MCP server and the on-device SDK for launch.
If you build voice agents, run a contact centre or make devices that listen, and want early access, a pilot or a partnership, get in touch.
See Formanta on your own audio
Ask for early access to the API and the MCP server, the SDK beta, an on-premise deployment or a pilot on recordings from your environment.