Data-driven speech processing models usually perform well with a large amount of text supervision, but collecting transcribed speech data is costly. Therefore, we propose SpeechCLIP, a novel framework bridging speech and text through images to enhance speech models without transcriptions. We leverage state-of-the-art pre-trained HuBERT and CLIP, aligning them via paired images and spoken captions with minimal fine-tuning. SpeechCLIP outperforms prior state-of-the-art on image-speech retrieval and performs zero-shot speech-text retrieval without direct supervision from transcriptions. Moreover, SpeechCLIP can directly retrieve semantically related keywords from speech.
Connectionist Temporal Classification (CTC) is a widely used approach for automatic speech recognition (ASR) that performs conditionally independent monotonic alignment. However for translation, CTC exhibits clear limitations due to the contextual and non-monotonic nature of the task and thus lags behind attentional decoder approaches in terms of translation quality. In this work, we argue that CTC does in fact make sense for translation if applied in a joint CTC/attention framework wherein CTC's core properties can counteract several key weaknesses of pure-attention models during training and decoding. To validate this conjecture, we modify the Hybrid CTC/Attention model originally proposed for ASR to support text-to-text translation (MT) and speech-to-text translation (ST). Our proposed joint CTC/attention models outperform pure-attention baselines across six benchmark translation tasks.
We introduce a framework for audio source separation using embeddings on a hyperbolic manifold that compactly represent the hierarchical relationship between sound sources and time-frequency features. Inspired by recent successes modeling hierarchical relationships in text and images with hyperbolic embeddings, our algorithm obtains a hyperbolic embedding for each time-frequency bin of a mixture signal and estimates masks using hyperbolic softmax layers. On a synthetic dataset containing mixtures of multiple people talking and musical instruments playing, our hyperbolic model performed comparably to a Euclidean baseline in terms of source to distortion ratio, with stronger performance at low embedding dimensions. Furthermore, we find that time-frequency regions containing multiple overlapping sources are embedded towards the center (i.e., the most uncertain region) of the hyperbolic space, and we can use this certainty estimate to efficiently trade-off between artifact introduction and interference reduction when isolating individual sounds.
HTR models development has become a conventional step for digital humanities projects. The performance of these models, often quite high, relies on manual transcription and numerous handwritten documents. Although the method has proven successful for Latin scripts, a similar amount of data is not yet achievable for scripts considered poorly-endowed, like Arabic scripts. In that respect, we are introducing and assessing a new modus operandi for HTR models development and fine-tuning dedicated to the Arabic Maghrib{\=i} scripts. The comparison between several state-of-the-art HTR demonstrates the relevance of a word-based neural approach specialized for Arabic, capable to achieve an error rate below 5% with only 10 pages manually transcribed. These results open new perspectives for Arabic scripts processing and more generally for poorly-endowed languages processing. This research is part of the development of RASAM dataset in partnership with the GIS MOMM and the BULAC.
Building systems that achieve a deeper understanding of language is one of the central goals of natural language processing (NLP). Towards this goal, recent works have begun to train language models on narrative datasets which require extracting the most critical information by integrating across long contexts. However, it is still an open question whether these models are learning a deeper understanding of the text, or if the models are simply learning a heuristic to complete the task. This work investigates this further by turning to the one language processing system that truly understands complex language: the human brain. We show that training language models for deeper narrative understanding results in richer representations that have improved alignment to human brain activity. We further find that the improvements in brain alignment are larger for character names than for other discourse features, which indicates that these models are learning important narrative elements. Taken together, these results suggest that this type of training can indeed lead to deeper language understanding. These findings have consequences both for cognitive neuroscience by revealing some of the significant factors behind brain-NLP alignment, and for NLP by highlighting that understanding of long-range context can be improved beyond language modeling.
The widely studied task of Natural Language Inference (NLI) requires a system to recognize whether one piece of text is textually entailed by another, i.e. whether the entirety of its meaning can be inferred from the other. In current NLI datasets and models, textual entailment relations are typically defined on the sentence- or paragraph-level. However, even a simple sentence often contains multiple propositions, i.e. distinct units of meaning conveyed by the sentence. As these propositions can carry different truth values in the context of a given premise, we argue for the need to recognize the textual entailment relation of each proposition in a sentence individually. We propose PropSegmEnt, a corpus of over 35K propositions annotated by expert human raters. Our dataset structure resembles the tasks of (1) segmenting sentences within a document to the set of propositions, and (2) classifying the entailment relation of each proposition with respect to a different yet topically-aligned document, i.e. documents describing the same event or entity. We establish strong baselines for the segmentation and entailment tasks. Through case studies on summary hallucination detection and document-level NLI, we demonstrate that our conceptual framework is potentially useful for understanding and explaining the compositionality of NLI labels.
Prior works on improving speech quality with visual input typically study each type of auditory distortion separately (e.g., separation, inpainting, video-to-speech) and present tailored algorithms. This paper proposes to unify these subjects and study Generalized Speech Enhancement, where the goal is not to reconstruct the exact reference clean signal, but to focus on improving certain aspects of speech. In particular, this paper concerns intelligibility, quality, and video synchronization. We cast the problem as audio-visual speech resynthesis, which is composed of two steps: pseudo audio-visual speech recognition (P-AVSR) and pseudo text-to-speech synthesis (P-TTS). P-AVSR and P-TTS are connected by discrete units derived from a self-supervised speech model. Moreover, we utilize self-supervised audio-visual speech model to initialize P-AVSR. The proposed model is coined ReVISE. ReVISE is the first high-quality model for in-the-wild video-to-speech synthesis and achieves superior performance on all LRS3 audio-visual enhancement tasks with a single model. To demonstrates its applicability in the real world, ReVISE is also evaluated on EasyCom, an audio-visual benchmark collected under challenging acoustic conditions with only 1.6 hours of training data. Similarly, ReVISE greatly suppresses noise and improves quality. Project page: https://wnhsu.github.io/ReVISE.
When using text data, social scientists often classify documents in order to use the resulting document labels as an outcome or predictor. Since it is prohibitively costly to label a large number of documents manually, automated text classification has become a standard tool. However, current approaches for text classification do not take advantage of all the data at one's disposal. We propose a fast new model for text classification that combines information from both labeled and unlabeled data with an active learning component, where a human iteratively labels documents that the algorithm is least certain about. Using text data from Wikipedia discussion pages, BBC News articles, historical US Supreme Court opinions, and human rights abuse allegations, we show that by introducing information about the structure of unlabeled data and iteratively labeling uncertain documents, our model improves performance relative to classifiers that (a) only use information from labeled data and (b) randomly decide which documents to label at the cost of manually labelling a small number of documents.
In this paper we present new attacks against federated learning when used to train natural language text models. We illustrate the effectiveness of the attacks against the next word prediction model used in Google's GBoard app, a widely used mobile keyboard app that has been an early adopter of federated learning for production use. We demonstrate that the words a user types on their mobile handset, e.g. when sending text messages, can be recovered with high accuracy under a wide range of conditions and that counter-measures such a use of mini-batches and adding local noise are ineffective. We also show that the word order (and so the actual sentences typed) can be reconstructed with high fidelity. This raises obvious privacy concerns, particularly since GBoard is in production use.
With the rapid development of ICT Custom Services (ICT CS) in power industries, the deployed power ICT CS systems mainly rely on the experience of customer service staff for fault type recognition, questioning, and answering, which makes it difficult and inefficient to precisely resolve the problems issued by users. To resolve this problem, in this paper, firstly, a multi-label fault text classification ensemble approach called BR-GBDT is proposed by combining Binary Relevance and Gradient Boosting Decision Tree for assisted fault type diagnosis and improving the accuracy of fault type recognition. Second, for the problem that there is lack of the training set for power ICT multi-label text classification, an automatic approach is presented to construct the training set from the historical fault text data stored in power ICT CS systems. The extensive experiments were made based on the power ICT CS training set and some general-purpose benchmark training datasets. The experiment results show that our approach outperforms the well known ensemble learning based approaches BR+LR and ML-KNN for fault text classification, efficiently handling the multi-label classification of ICT custom service text data for fault type recognition.