On The Performance of Time-Pooling Strategies for End-to-End Spoken Language Identification

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3566–3572 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3566 On The Performance of Time-Pooling Strategies for End-to-End Spoken Language Identification Jo˜ao Monteiro1,2, Jahangir Alam1,2, Tiago H. Falk1 1Institut National de la Recherche Scientifique (INRS-EMT), Quebec, Canada 2Centre de Recherche Informatique de Montr´eal (CRIM), Quebec, Canada joao.monteiro@emt.inrs.ca, jahangir.alam@crim.ca, falk@emt.inrs.ca Abstract Automatic speech processing applications often have to deal with the problem of aggregating local descriptors (i.e., representations of input speech data corresponding to specific portions across the time dimension) and turning them into a single fixed-dimension representation, known as global descriptor, on top of which downstream classification tasks can be performed. In this paper, we provide an empirical assessment of different time pooling strategies when used with state-of-the-art representation learning models. In particular, insights are provided as to when it is suitable to use simple statistics of local descriptors or when more sophisticated approaches are needed. Here, language identification is used as a case study and a database containing ten oriental languages under varying test conditions (short-duration test recordings, confusing languages, unseen languages) is used. Experiments are performed with classifiers trained on top of global descriptors to provide insights on open-set evaluation performance and show that appropriate selection of such pooling strategies yield embeddings able to outperform well-known benchmark systems as well as previously results based on attention only. Keywords: Spoken language identification, Temporal pooling, End-to-end identification Feature Extraction Representation Model V1 V2 V3 VN . . . Input signal Local descriptors Temporal Pooling V Global descriptor Figure 1: Mapping an input signal to a low-dimensional space. Pooling across the time dimension combines a set of local descriptors into a single global descriptor vector. 1. Introduction Language identification (LID) from speech corresponds to the task of identifying the spoken language from a given speech example under the assumption that a single lan- guage is present. LID is commonly tackled using similar approaches as those employed for e.g., speaker verifica- tion/recognition. Representative examples include the use of low-dimensional language-dependent i-vectors (Dehak et al., 2011b), which are similar to speaker-dependent i- vectors used in speaker recognition (Dehak et al., 2011a). These are obtained by first computing a generative model of frame-level features, i.e., a universal background model (UBM), followed by factor analysis performed on top of statistics of latent variables in order to obtain a fixed- dimension representation that is independent of the input speech duration. Downstream tasks can then rely on these representations rather than general-purpose features. In recent years, following the success of deep learning, generative modeling (Goodfellow et al., 2014) and deep neural networks have become a popular alternative to i- vectors. Commonly, such methods are employed to ei- ther map features obtained from speech recordings directly to low-dimensional representations (Zhong et al., 2017) or to perform recognition in an end-to-end fashion (Rohdin et al., 2017; Snyder et al., 2016), thus bypassing a sec- ond classification stage. Representative examples include the …

Content truncated to 500 words