Voice cloning, an emerging field in the speech processing area, aims to generate synthetic utterances that closely resemble the voices of specific individuals. In this study, we investigate the impact of various techniques on improving the quality of voice cloning, specifically focusing on a low-quality dataset. To contrast our findings, we also use two high-quality corpora for comparative analysis. We conduct exhaustive evaluations of the quality of the gathered corpora in order to select the most suitable audios for the training of a Voice Cloning system. Following these measurements, we conduct a series of ablations by removing audios with lower SNR and higher variability in utterance speed from the corpora in order to decrease their heterogeneity. Furthermore, we introduce a novel algorithm that calculates the fraction of aligned input characters by exploiting the attention matrix of the Tacotron 2 Text-to-Speech (TTS) system. This algorithm provides a valuable metric for evaluating the alignment quality during the voice cloning process. We present the results of our experiments, demonstrating that the performed ablations significantly increase the quality of synthesised audios for the challenging low-quality corpus. Notably, our findings indicate that models trained on a 3-hour corpus from a pre-trained model exhibit comparable audio quality to models trained from scratch using significantly larger amounts of data.

The growth in online child exploitation material is a significant challenge for European Law Enforcement Agencies (LEAs). One of the most important sources of such online information corresponds to audio material that needs to be analyzed to find evidence in a timely and practical manner. That is why LEAs require a next-generation AI-powered platform to process audio data from online sources. We propose the use of speech recognition and keyword spotting to transcribe audiovisual data and to detect the presence of keywords related to child abuse. The considered models are based on two of the most accurate neural-based architectures to date: Wav2vec2.0 and Whisper. The systems are tested under an extensive set of scenarios in different languages. Additionally, keeping in mind that obtaining data from LEAs is very sensitive, we explore the use of federated learning to have more robust systems for the addressed application, while maintaining the privacy of the data to LEAs. The considered models achieved a word error rate between 11% and 25%, depending on the language. In addition, the systems are able to recognize a set of spotted words with true positives rates between 82% and 98%, depending on the language. Finally, federated learning strategies show that they can maintain and even improve the performance of the systems when compared to centralized trained models. The proposed systems sit the basis for an AI-powered platform for automatic analysis of audio in the context of forensic applications within child abuse. The use of federated learning is also promising for the addressed scenario, where data privacy is an important issue to be managed.