State of the art diarization systems now achieve decent performance but those performances are often not good enough to deploy them without any human supervision. Additionally, most approaches focus on single audio files while many use cases involving multiple recordings with recurrent speakers require the incremental processing of a collection. In this paper, we propose a framework that solicits a human in the loop to correct the clustering by answering simple questions. After defining the nature of the questions for both single file and collection of files, we propose two algorithms to list those questions and associated stopping criteria that are necessary to limit the work load on the human in the loop. Experiments performed on the ALLIES dataset show that a limited interaction with a human expert can lead to considerable improvement of up to 36.5% relative diarization error rate (DER) for single files and 33.29% for a collection.

In affective computing, two main paradigms are used to represent emotion: categorical representation and dimensional description in continuous space. Therefore, the emotion recognition task can be treated as a classification or regression. The main aim of this study is to investigate the relation between these two representations and propose a classification pipeline that uses only dimensional annotation. Our approach contains a neural regressor which predicts a vector of arousal, valence and dominance values for a given speech segment. This vector can be interpreted as an emotional category using a mapping algorithm. We investigate the performances of a neural network architectures, and three mapping algorithms on two corpora. Our study shows the limitation and an advantage of the emotion classification via regression approach.

Subjective listening tests are the standard for evaluating synthetic speech quality, but they are costly, time-consuming, and difficult to scale. Automatic speech synthetic quality predictors reduce reliance on human ratings, but still depend on large amounts of human-annotated data and often poorly generalize across languages, speakers, and synthesis setups. This paper proposes a humanless annotation framework for learning synthetic speech quality predictors by exploiting training dynamics of text-to-speech systems as pseudo-annotations. Synthetic speech is collected from multiple Text-To-Speech (TTS) models at different training checkpoints and paired with training-derived metadata such as iteration number or loss. A quality predictor is then directly trained from audio without human perceptual labels. Experiments conducted on multiple benchmark datasets covering different languages, speakers, domains, and TTS systems show that checkpoint iteration provides a reliable proxy for perceptual quality and correlates with human ratings across evaluation conditions. Results further demonstrate strong generalization to unseen synthesis setups and languages, including a low-resource Kurdish setting. We also investigate the role of linguistic diversity in the training data and show that increased linguistic variation improves the robustness and transferability of the learned predictor. These findings suggest that TTS training dynamics constitute an effective source of scalable supervision for speech quality prediction without human annotations.