Speaker Identification and Verification
Recognizing who is speaking from their voice: verifying a claimed identity, picking a speaker from a known group, or splitting a recording by speaker.
Also speaker recognition · voice biometrics · speaker verification · speaker diarization · voiceprint · x-vectors
You can tell a friend's voice on the phone before they say their name. Speaker recognition teaches a computer the same trick, using the shape and texture of each person's voice.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Uses / built on
How it works
Speech recognition asks what was said. Speaker recognition asks who said it.
Three related tasks:
- Verification: “Is this really Alice?” Compare the voice with Alice’s enrolled sample.
- Identification: “Which of our 50 employees is this?”
- Diarization: “Who spoke when?” in a recording with several people.
Modern systems all rely on speaker embeddings: a neural network listens to a few seconds of audio (as a spectrogram) and outputs a vector that captures the voice, not the words. Recordings of the same person land close together in this embedding space; deciding “same speaker or not” becomes measuring a distance, the same idea as in face recognition. Before deep learning, voices were modeled with Gaussian mixture models.
The big caveat today: text-to-speech voice cloning can imitate a person from a short sample, so banks and others that relied on “my voice is my password” have had to add stronger checks.
Everyday use cases
- Case 01
Personalized smart speakers
A home assistant recognizing which family member is asking, to read out the right calendar.
- Case 02
Meeting transcripts
Labeling "Speaker 1" and "Speaker 2" in a recorded call (diarization).
Strengths, weaknesses and when to use it
Strengths
- Works hands-free and remotely, with only a few seconds of audio.
Weaknesses
- Voice cloning makes spoofing much easier, so voice alone is now a weak password.
- Colds, noise and microphones change how a voice sounds.
When to use itPersonalization and transcripts; for security, only combined with other factors and anti-spoofing checks.
Compare with
Check yourself
- Q01
What does speaker diarization answer?
Show answer
A: Who spoke when