Effects of language mismatch in automatic forensic voice comparison using deep learning embeddings

TL;DR

This study investigates the impact of language mismatch on deep learning-based forensic voice comparison, showing that models trained on large multilingual datasets perform well even on low-resource languages like Hungarian.

Contribution

It demonstrates that pre-trained speaker embedding models on large multilingual datasets can effectively be used for forensic voice comparison in low-resource languages, despite language differences.

Findings

Models trained on large multilingual datasets perform well across languages.

Sample duration positively influences verification performance.

Speaking style variations have minimal impact on results.

Abstract

In forensic voice comparison the speaker embedding has become widely popular in the last 10 years. Most of the pretrained speaker embeddings are trained on English corpora, because it is easily accessible. Thus, language dependency can be an important factor in automatic forensic voice comparison, especially when the target language is linguistically very different. There are numerous commercial systems available, but their models are mainly trained on a different language (mostly English) than the target language. In the case of a low-resource language, developing a corpus for forensic purposes containing enough speakers to train deep learning models is costly. This study aims to investigate whether a model pre-trained on English corpus can be used on a target low-resource language (here, Hungarian), different from the model is trained on. Also, often multiple samples are not available from the offender (unknown speaker). Therefore, samples are compared pairwise with and without speaker enrollment for suspect (known) speakers. Two corpora are applied that were developed especially for forensic purposes, and a third that is meant for traditional speaker verification. Two deep learning based speaker embedding vector extraction methods are used: the x-vector and ECAPA-TDNN. Speaker verification was evaluated in the likelihood-ratio framework. A comparison is made between the language combinations (modeling, LR calibration, evaluation). The results were evaluated by minCllr and EER metrics. It was found that the model pre-trained on a different language but on a corpus with a huge amount of speakers performs well on samples with language mismatch. The effect of sample durations and speaking styles were also examined. It was found that the longer the duration of the sample in question the better the performance is. Also, there is no real difference if various speaking styles are applied.