Text-dependent Forensic Voice Comparison: Likelihood Ratio Estimation with the Hidden Markov Model (HMM) and Gaussian Mixture Model

Text-dependent Forensic Voice Comparison: Likelihood Ratio Estima- tion with the Hidden Markov Model (HMM) and Gaussian Mixture Model – Universal Background Model (GMM-UBM) Approaches Satoru Tsuge Daido University, Japan tsuge@daido-it.ac.jp Shunichi Ishihara Australian National University shunichi.ishihara@anu.edu.au Abstract Among the more typical forensic voice comparison (FVC) approaches, the acous- tic-phonetic statistical approach is suitable for text-dependent FVC, but it does not fully exploit available time-varying infor- mation of speech in its modelling. The au- tomatic approach, on the other hand, es- sentially deals with text-independent cas- es, which means temporal information is not explicitly incorporated in the model- ling. Text-dependent likelihood ratio (LR)- based FVC studies, in particular those that adopt the automatic approach, are few. This preliminary LR-based FVC study compares two statistical models, the Hid- den Markov Model (HMM) and the Gaussian Mixture Model (GMM), for the calculation of forensic LRs using the same speech data. FVC experiments were car- ried out using different lengths of Japanese short words under a forensically realistic, but challenging condition: only two speech tokens for model training and LR estima- tion. Log-likelihood-ratio cost (Cllr) was used as the assessment metric. The study demonstrates that the HMM system con- stantly outperforms the GMM system in terms of average Cllr values. However, words longer than three mora are needed if the advantage of the HMM is to become evident. With a seven-mora word, for ex- ample, the HMM outperformed the GMM by a Cllr value of 0.073. 1 Introduction After the DNA success story, the likelihood ratio (LR)-based approach became the new paradigm for evaluating and presenting forensic evidence in court. The LR approach has also been applied to speech evidence(Rose, 2006), and it is increasing- ly accepted in forensic voice comparison (FVC) as well (Morrison, 2009). There are two different approaches in FVC. They are the ‘acoustic-phonetic statistical ap- proach’ and the ‘automatic approach’ (Morrison et al., 2018). The former usually works on compara- ble phonetic units that can be found in both the of- fender and suspect samples. In the latter, acoustic measurements are usually carried out over all por- tions of the available recordings, resulting in more detailed acoustic characteristics of the speakers. The common statistical models used in the auto- matic approach are the Gaussian mixture model – universal background model (GMM-UBM) (Reynolds et al., 2000) and i-vectors with proba- bilistic linear discrimination analysis (PLDA) (Burget et al., 2011). Due to its nature, the auto- matic approach is mainly used for text- ‘independent’ FVC, and there is a good amount of research on this (Enzinger & Morrison, 2017; Enzinger et al., 2016). The acoustic-phonetic sta- tistical approach is a type of text-‘dependent’ FVC because it tends to focus on particular linguistic units, such as phonemes, words, phrases, etc. Hav- ing said that, even if one is targeting a particular word or phrase, for example ‘hello’, all obtainable features are not exploited in the acoustic-phonetic statistical approach because it still tends to focus on particular segments or phonemes of the word or …

Content truncated to 500 words