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 were tested under an extensive set of scenarios in different languages. Additionally, keeping in mind that obtaining data from LEAs are very sensitive, we explore the use of federated learning to provide more robust systems for the addressed application, while maintaining the privacy of the data from 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-positive 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 set the basis for an AI-powered platform for automatic analysis of audio in the context of forensic applications of child abuse. The use of federated learning is also promising for the addressed scenario, where data privacy is an important issue to be managed.
Automatic speech recognition in patients with aphasia is a challenging task for which studies have been published in a few languages. Reasonably, the systems reported in the literature within this field show significantly lower performance than those focused on transcribing non-pathological clean speech. It is mainly due to the difficulty of recognizing a more unintelligible voice, as well as due to the scarcity of annotated aphasic data. This work is mainly focused on applying novel semi-supervised learning methods to the AphasiaBank dataset in order to deal with these two major issues, reporting improvements for the English language and providing the first benchmark for the Spanish language for which less than one hour of transcribed aphasic speech was used for training. In addition, the influence of reinforcing the training and decoding processes with out-of-domain acoustic and text data is described by using different strategies and configurations to fine-tune the hyperparameters and the final recognition systems. The interesting results obtained encourage extending this technological approach to other languages and scenarios where the scarcity of annotated data to train recognition models is a challenging reality.
Speech translation has been traditionally tackled under a cascade approach, chaining speech recognition and machine translation components to translate from an audio source in a given language into text or speech in a target language. Leveraging on deep learning approaches to natural language processing, recent studies have explored the potential of direct end-to-end neural modelling to perform the speech translation task. Though several benefits may come from end-to-end modelling, such as a reduction in latency and error propagation, the comparative merits of each approach still deserve detailed evaluations and analyses. In this work, we compared state-of-the-art cascade and direct approaches on the under-resourced Basque–Spanish language pair, which features challenging phenomena such as marked differences in morphology and word order. This case study thus complements other studies in the field, which mostly revolve around the English language. We describe and analysed in detail the mintzai-ST corpus, prepared from the sessions of the Basque Parliament, and evaluated the strengths and limitations of cascade and direct speech translation models trained on this corpus, with variants exploiting additional data as well. Our results indicated that, despite significant progress with end-to-end models, which may outperform alternatives in some cases in terms of automated metrics, a cascade approach proved optimal overall in our experiments and manual evaluations.
In this paper, a new supervised classification paradigm, called classifier subset selection for stacked generalization (CSS stacking), is presented to deal with speech emotion recognition. The new approach consists of an improvement of a bi-level multi-classifier system known as stacking generalization by means of an integration of an estimation of distribution algorithm (EDA) in the first layer to select the optimal subset from the standard base classifiers. The good performance of the proposed new paradigm was demonstrated over different configurations and datasets. First, several CSS stacking classifiers were constructed on the RekEmozio dataset, using some specific standard base classifiers and a total of 123 spectral, quality and prosodic features computed using in-house feature extraction algorithms. These initial CSS stacking classifiers were compared to other multi-classifier systems and the employed standard classifiers built on the same set of speech features. Then, new CSS stacking classifiers were built on RekEmozio using a different set of both acoustic parameters (extended version of the Geneva Minimalistic Acoustic Parameter Set (eGeMAPS)) and standard classifiers and employing the best meta-classifier of the initial experiments. The performance of these two CSS stacking classifiers was evaluated and compared. Finally, the new paradigm was tested on the well-known Berlin Emotional Speech database. We compared the performance of single, standard stacking and CSS stacking systems using the same parametrization of the second phase. All of the classifications were performed at the categorical level, including the six primary emotions plus the neutral one.
Study of emotions in human–computer interaction is a growing research area. This paper shows an attempt to select the most significant features for emotion recognition in spoken Basque and Spanish Languages using different methods for feature selection. RekEmozio database was used as the experimental data set. Several Machine Learning paradigms were used for the emotion classification task. Experiments were executed in three phases, using different sets of features as classification variables in each phase. Moreover, feature subset selection was applied at each phase in order to seek for the most relevant feature subset. The three phases approach was selected to check the validity of the proposed approach. Achieved results show that an instance-based learning algorithm using feature subset selection techniques based on evolutionary algorithms is the best Machine Learning paradigm in automatic emotion recognition, with all different feature sets, obtaining a mean of 80,05% emotion recognition rate in Basque and a 74,82% in Spanish. In order to check the goodness of the proposed process, a greedy searching approach (FSS-Forward) has been applied and a comparison between them is provided. Based on achieved results, a set of most relevant non-speaker dependent features is proposed for both languages and new perspectives are suggested.