- high performance necessary under 'real world' conditions
- difficulties
- intrinsic: age, emotion, manner of speaking
- extrinsic: background noise, reverberation, channel/mic
- speaker identification: mapping a given utterance to a speaker (open set vs. closed set)
- makes "closed set" a "multi-class classification"
- classification loss
- speaker verification: mapping a given utterance to a target model
- contrastive loss (learn the embedding, rather than computing distance, e.g. Siamese)
- a portion of data (=test set) left for unseen POIs
- practice
- model
- average pooling
- fixed input length
- metric
- EER
- terms:
- POI: person of interest
Speech Processing/AI/Linguistics/CS/etc.