In this paper, we study how word-like units are represented and activated in a recurrent neural model of visually grounded speech. The model used in our experiments is trained to project an image and its spoken description in a common representation space. We show that a recurrent model trained on spoken sentences implicitly segments its input into word-like units and reliably maps them to their correct visual referents. We introduce a methodology originating from linguistics to analyse the representation learned by neural networks -the gating paradigm -and show that the correct representation of a word is only activated if the network has access to first phoneme of the target word, suggesting that the network does not rely on a global acoustic pattern. Furthermore, we find out that not all speech frames (MFCC vectors in our case) play an equal role in the final encoded representation of a given word, but that some frames have a crucial effect on it. Finally, we suggest that word representation could be activated through a process of lexical competition. Related WorkIn this section we explore what is known about word recognition in humans. We then review recent works related to the representation of lan-arXiv:1909.08491v1 [cs.CL]

We investigate the behaviour of attention in neural models of visually grounded speech trained on two languages: English and Japanese. Experimental results show that attention focuses on nouns and this behaviour holds true for two very typologically different languages. We also draw parallels between artificial neural attention and human attention and show that neural attention focuses on word endings as it has been theorised for human attention. Finally, we investigate how two visually grounded monolingual models can be used to perform cross-lingual speech-to-speech retrieval. For both languages, the enriched bilingual (speech-image) corpora with part-of-speech tags and forced alignments are distributed to the community for reproducible research.Index Termsgrounded language learning, attention mechanism, cross-lingual speech retrieval, recurrent neural networks.

This paper presents an augmentation of MSCOCO dataset where speech is added to image and text. Speech captions are generated using text-to-speech (TTS) synthesis resulting in 616,767 spoken captions (more than 600h) paired with images. Disfluencies and speed perturbation are added to the signal in order to sound more natural. Each speech signal (WAV) is paired with a JSON file containing exact timecode for each word/syllable/phoneme in the spoken caption. Such a corpus could be used for Language and Vision (LaVi) tasks including speech input or output instead of text. Investigating multimodal learning schemes for unsupervised speech pattern discovery is also possible with this corpus, as demonstrated by a preliminary study conducted on a subset of the corpus (10h, 10k spoken captions).

The language acquisition literature shows that children do not build their lexicon by segmenting the spoken input into phonemes and then building up words from them, but rather adopt a top-down approach and start by segmenting word-like units and then break them down into smaller units. This suggests that the ideal way of learning a language is by starting from full semantic units. In this paper, we investigate if this is also the case for a neural model of Visually Grounded Speech trained on a speech-image retrieval task. We evaluated how well such a network is able to learn a reliable speech-to-image mapping when provided with phone, syllable, or word boundary information. We present a simple way to introduce such information into an RNN-based model and investigate which type of boundary is the most efficient. We also explore at which level of the network's architecture such information should be introduced so as to maximise its performances. Finally, we show that using multiple boundary types at once in a hierarchical structure, by which low-level segments are used to recompose high-level segments, is beneficial and yields better results than using lowlevel or high-level segments in isolation.