Speaker intelligence API
No credit card required
Built by the team behind the most-used diarization model: pyannote.audio - 1B+ download.
In production in 500+ enterprises - Synthesia, Descript, SmallestAI, Notta, Gladia, UpMeet...
Identification result
4 voiceprints enrolled
Alice Chen
0.94
00:00 → 00:12
IDENTIFIED
James Okafor
0.81
00:12 → 00:45
IDENTIFIED
SPEAKER_02
0.58
00:45 → 00:58
BELOW THRESHOLD
SPEAKER_03
0.31
00:58 → 01:03
BELOW THRESHOLD
Matching threshold
0.62
Segments scoring below the threshold stay anonymous instead of taking a name.
Twelve years of diarization research
Voice Prints
Create a voiceprint from a short audio sample of a known speaker. A voiceprint is a fixed-size embedding of a voice, not a comparison, so enrollment happens once and is reused across every later request.
Speaker Identification
Match segments of new audio against your enrolled voiceprints. The API returns a resolved identity per segment and a confidence score per candidate, so your application decides what to accept.
Run alongside Diarization: the same pipeline can label and then name speakers.
Enroll
Send a short audio sample of a known speaker and store the voiceprint returned to you.
Match services compliance
Send new audio, and each speaker segment is compared against the pool of enrolled voiceprints.
Score
Every candidate comes back with a confidence score, and a matching threshold controls the minimum score required before a name is applied.
Match quality depends on three things: how much clean speech the enrollment sample contains, the acoustic conditions of the audio being analyzed, and channel mismatch between enrollment and inference, such as a headset voiceprint matched against a phone recording.
In production today
Contact centers
Label agent and customer automatically, and keep the same identity across every call in the account history.
Meeting intelligence
Name participants instead of numbering them, attribute action items to a specific name, and power role-aware analytics such as talk-time per attendee.
Media archives
Find a voice across a catalog, credit podcast hosts and recurring guests, and build interview archives that are searchable by person.
Moderation and compliance
Flag which speaker said what, measure talk time by speaker, and keep an attributed record for audit.
Voice AI training
Clean up multi-speaker transcripts before they become training or evaluation data.
Voiceprints are biometric data. Speaker identification works by comparing voice signatures, and that carries obligations that ordinary transcription does not.
We do not store your voiceprints
The API returns the voiceprint to you as an encrypted value that you hold in your own system. Raw embedding values are not exposed.
Consent belongs upstream
If your product identifies people by voice, the people being identified need to have agreed to it. Build the consent step into your enrollment flow, not around it.
Deployment follows your data
The same models run in our cloud API or inside your own infrastructure through the on-premises package, so regulated audio can stay where it has to stay.
This is factual guidance, not legal advice. Retention periods, lawful basis, and disclosure requirements depend on your jurisdiction and your users.
Speaker recognition is the task of recognizing a person from the characteristics of their voice.
Unlike diarization, recognition requires a reference: a database of enrolled voice profiles against which incoming audio is compared. The vocabulary around it is used loosely, so it helps to separate the terms:
Speaker recognition is the umbrella term for techniques that match a voice to a known identity.
Speaker identification answers "which known speaker is this?" by comparing a voice against a pool of enrolled profiles.
Speaker verification answers "is this the person they claim to be?" by comparing a voice against one specific profile.
Speaker diarization answers "who spoke when?" and leaves the speakers unnamed.
A voiceprint is a compact voice fingerprint created from audio: a fixed-size embedding capturing the acoustic characteristics of a speaker.
Creating a voiceprint involves no comparison at all. It is simply a representation of a voice, generated from one or more audio samples. The comparison is what turns it into a capability, and how you compare determines which capability you get.
Diarization is a general-purpose, enrollment-free capability that works on any audio, with any speakers, with no prior setup. It segments the timeline and returns anonymous labels such as SPEAKER_00.
Identification is for known voices. Given a set of enrolled voiceprints, it resolves those anonymous labels into real identities:
Alice Chen (Customer): 00:00 → 00:12 "We're having trouble with onboarding..."
James Okafor (AE): 00:12 → 00:45 "I understand, let me walk you through..."
Alice Chen (Customer): 00:45 → 01:03 "That makes more sense now, thank you."
The two chain naturally: diarize first, then identify. See the Speaker Diarization API page for the diarization layer on its own.
If you are moving off a retired service, three steps cover most of the work:
01
Re-enroll your voiceprints. Voice profiles are not portable between providers. You need the original enrollment audio, or a new enrollment pass with your users.
02
Map your verification and identification calls. One-to-one checks map to verification, one-to-many lookups map to identification. Identification also returns the diarized timeline, so a separate segmentation call may no longer be needed.
03
Revalidate thresholds on your own audio. Confidence scores are not comparable across systems. Re-derive your matching threshold from a labeled sample of your own recordings before you go live.
What is the difference between speaker recognition and speaker identification?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.
How much audio does a voiceprint need?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.
How to add speaker identification into a transcription pipeline?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.
What replaces other speaker recognition services?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.