Speaker Verification Embeddings
Mapping a voice to a fixed-length embedding to verify or identify the speaker.
Tier Grades
Tier A: Frontier AI labs
feasible
NVIDIA and Alibaba publish speaker embedding models.
Tier B: Indonesian AI labs
feasible
Models of 6M parameters trained on public VoxCeleb data are cheap for a national lab.
Tier C: Institution units such as ITB
feasible
ECAPA-TDNN came from a 3-author team on public data.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Released embeddings support speaker identification in a pilot, with spoofing risk noted.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeTitaNet-S has only 6M parameters; CAM++ targets ECAPA-TDNN accuracy at lower computational cost.
- DataVoxCeleb (ECAPA-TDNN, TitaNet EER 0.68%) and CN-Celeb (CAM++).
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 3 authors on ECAPA-TDNN; 5 on CAM++.
- Time horizonMonths.
- Shelf-life signalsVoice cloning defeating voice biometrics, shifting work to spoofing detection
- Pipeline roleTier A and Tier C release embedding models; Tier B tests them on Indonesian voices; Tier D uses them inside diarization and caller verification pilots.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Links
R&D TopicSpeaker Verification Embeddings
Models
CAM++Alibaba (Qwen), speaker
ECAPA-TDNN (SpeechBrain)SpeechBrain community, speaker
TitaNetNVIDIA, speakerERes2NetAlibaba (Qwen), speaker+1 more
CAM++Alibaba (Qwen), speakerECAPA-TDNN (SpeechBrain)SpeechBrain community, speakerTitaNetNVIDIA, speakerERes2NetAlibaba (Qwen), speaker+1 moreUse Casesnone recorded
Who Runs Itnone recorded
Papers
- Generalized End-to-End Loss for Speaker VerificationLi Wan et al., ICASSP 2018 (per arXiv comment), 2017 founding
- VoxCeleb: a large-scale speaker identification datasetArsha Nagrani et al., arXiv preprint, 2017 founding
- CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware MaskingHui Wang et al., arXiv preprint, 2023 state of the art
- ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker VerificationBrecht Desplanques et al., INTERSPEECH 2020 (per arXiv comment), 2020 state of the art
- In defence of metric learning for speaker recognitionJoon Son Chung et al., arXiv preprint, 2020 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2005.07143 https://arxiv.org/abs/2005.07143
- arxiv.org/abs/2110.04410 https://arxiv.org/abs/2110.04410
- arxiv.org/abs/2303.00332 https://arxiv.org/abs/2303.00332
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Speaker Verification Embeddings", https://aurion.semai.id/topics/speaker-verification-embeddings/, records of October 2026.