Topic-based speaker recognition for German parliamentary speeches

In the last decade, high-level features for speaker recognition have become a research focus, as they are believed to alleviate the weak point of the classical spectral/cepstral-feature-based approaches: mismatch in acoustic conditions or channel between training and test data. Identification cues such as prosody, pronunciation, and idiolect have been successfully investigated. Semantic speaker recognition, such as identifying people by the topics they frequently talk about, has not found an equal amount of attention. However, it is a promising approach, especially for broadcast data and multimedia archives, where prominent speakers can be expected to often talk about their specific subjects. This paper reports on our experiments with topic-based speaker recognition on German parliamentary speeches. Text transcripts of speeches of federal ministers were used to train speaker models based on word frequencies. For recognition, these models were applied to automatic speech recognition transcripts of parliamentary speeches and could identify the correct speaker surprisingly well, with an EER of 13.8%. Fusing this approach with a classical GMM-UBM system (with EER 14.3%) yields an improved EER of 8.6%.