Glossary
A speaker embedding is a fixed-length vector (typically 192 or 256 dimensions) that numerically represents the voice characteristics of a speaker — pitch, formants, vocal-tract resonance, speaking style. The same speaker produces similar embeddings across different recordings; different speakers produce different embeddings. Speaker embeddings are the engine behind speaker diarization and speaker verification.
How embeddings are computed
A neural network trained on thousands of hours of labeled speaker data (VoxCeleb, etc.) learns to output a vector for each input audio segment such that vectors from the same speaker cluster together and vectors from different speakers are far apart. Common architectures: x-vectors (TDNN-based), ECAPA-TDNN (current SOTA), WavLM speaker embeddings (transformer-based). Output dimensions: 192 (ECAPA) or 256 (x-vector).
Use case 1: diarization
Take a multi-speaker recording, segment it via VAD, compute a speaker embedding per segment, then cluster the embeddings (k-means or agglomerative). Each cluster = one speaker. Label them Speaker 1, Speaker 2, etc. The user renames them after the fact. This is exactly what runs when you click "Speaker ID" on a Transcript.you transcript.
Use case 2: speaker verification
"Is this Sarah's voice?" Compute Sarah's reference embedding from a known recording. Compute the new recording's embedding. Compute cosine similarity. Above threshold → same speaker. This is how voice biometrics (banking voice login, Siri "Hey Siri" personalization) work. We don't currently do verification — diarization is enough for transcript clarity.
In practice
Drop a 4-person Zoom recording into Transcript.you and click Speaker ID. We slice by VAD, compute an ECAPA embedding for every utterance (~1 second each), then cluster the embeddings into 4 groups. Each utterance gets a Speaker N label. Find-and-replace to rename to actual people. Total processing time: ~10 seconds for a 1-hour call.
Related terms
Further reading
Try multi-speaker transcription
Get started — freeLast updated: September 20, 2026
Frequently Asked Questions
What does Speaker embedding mean in transcription?
A fixed-length vector that captures the voice characteristics of a speaker. Used in diarization to cluster segments by speaker identity without knowing names in advance.
Why does Speaker embedding matter when choosing a transcription workflow?
Speaker embedding affects how you understand transcript quality, timing, compatibility, or the technology behind speech-to-text results.
Where can I apply Speaker embedding on Transcript.you?
Use the related workflow at /interview-transcription when you want to see how this glossary concept connects to an actual transcription task.
Browse the full transcription glossary or read the complete guide to AI transcription.