Hugo Van hamme

Lies Bollens, Tom Francart, Hugo Van Hamme

The electroencephalogram (EEG) is a powerful method to understand how the brain processes speech. Linear models have recently been replaced for this purpose with deep neural networks and yield promising results. In related EEG classification fields, it is shown that explicitly modeling subject-invariant features improves generalization of models across subjects and benefits classification accuracy. In this work, we adapt factorized hierarchical variational autoencoders to exploit parallel EEG recordings of the same stimuli. We model EEG into two disentangled latent spaces. Subject accuracy reaches 98.96% and 1.60% on respectively the subject and content latent space, whereas binary content classification experiments reach an accuracy of 51.51% and 62.91% on respectively the subject and content latent space.

Steven Vander Eeckt, Hugo Van hamme

Adapting a trained Automatic Speech Recognition (ASR) model to new tasks results in catastrophic forgetting of old tasks, limiting the model's ability to learn continually and to be extended to new speakers, dialects, languages, etc. Focusing on End-to-End ASR, in this paper, we propose a simple yet effective method to overcome catastrophic forgetting: weight averaging. By simply taking the average of the previous and the adapted model, our method achieves high performance on both the old and new tasks. It can be further improved by introducing a knowledge distillation loss during the adaptation. We illustrate the effectiveness of our method on both monolingual and multilingual ASR. In both cases, our method strongly outperforms all baselines, even in its simplest form.

Pu Wang, Hugo Van hamme

End-to-end spoken language understanding (SLU) systems benefit from pretraining on large corpora, followed by fine-tuning on application-specific data. The resulting models are too large for on-edge applications. For instance, BERT-based systems contain over 110M parameters. Observing the model is overparameterized, we propose lean transformer structure where the dimension of the attention mechanism is automatically reduced using group sparsity. We propose a variant where the learned attention subspace is transferred to an attention bottleneck layer. In a low-resource setting and without pre-training, the resulting compact SLU model achieves accuracies competitive with pre-trained large models.

Jakob Poncelet, Hugo Van hamme

TV subtitles are a rich source of transcriptions of many types of speech, ranging from read speech in news reports to conversational and spontaneous speech in talk shows and soaps. However, subtitles are not verbatim (i.e. exact) transcriptions of speech, so they cannot be used directly to improve an Automatic Speech Recognition (ASR) model. We propose a multitask dual-decoder Transformer model that jointly performs ASR and automatic subtitling. The ASR decoder (possibly pre-trained) predicts the verbatim output and the subtitle decoder generates a subtitle, while sharing the encoder. The two decoders can be independent or connected. The model is trained to perform both tasks jointly, and is able to effectively use subtitle data. We show improvements on regular ASR and on spontaneous and conversational ASR by incorporating the additional subtitle decoder. The method does not require preprocessing (aligning, filtering, pseudo-labeling, ...) of the subtitles.

Quentin Meeus, Marie-Francine Moens, Hugo Van hamme

We explore the benefits that multitask learning offer to speech processing as we train models on dual objectives with automatic speech recognition and intent classification or sentiment classification. Our models, although being of modest size, show improvements over models trained end-to-end on intent classification. We compare different settings to find the optimal disposition of each task module compared to one another. Finally, we study the performance of the models in low-resource scenario by training the models with as few as one example per class. We show that multitask learning in these scenarios compete with a baseline model trained on text features and performs considerably better than a pipeline model. On sentiment classification, we match the performance of an end-to-end model with ten times as many parameters. We consider 4 tasks and 4 datasets in Dutch and English.

Quentin Meeus, Marie-Francine Moens, Hugo Van hamme

Most spoken language understanding systems use a pipeline approach composed of an automatic speech recognition interface and a natural language understanding module. This approach forces hard decisions when converting continuous inputs into discrete language symbols. Instead, we propose a representation model to encode speech in rich bidirectional encodings that can be used for downstream tasks such as intent prediction. The approach uses a masked language modelling objective to learn the representations, and thus benefits from both the left and right contexts. We show that the performance of the resulting encodings before fine-tuning is better than comparable models on multiple datasets, and that fine-tuning the top layers of the representation model improves the current state of the art on the Fluent Speech Command dataset, also in a low-data regime, when a limited amount of labelled data is used for training. Furthermore, we propose class attention as a spoken language understanding module, efficient both in terms of speed and number of parameters. Class attention can be used to visually explain the predictions of our model, which goes a long way in understanding how the model makes predictions. We perform experiments in English and in Dutch.

Pouya Mehralian, Melissa Farasyn, Anne Breitbarth et al.

Automatic Speech Recognition (ASR) in dialect-heavy settings remains challenging due to strong regional variation and limited labeled data. We propose GLoRIA, a parameter-efficient adaptation framework that leverages metadata (e.g., coordinates) to modulate low-rank updates in a pre-trained encoder. GLoRIA injects low-rank matrices into each feed-forward layer, with a gating MLP determining the non-negative contribution of each LoRA rank-1 component based on location metadata. On the GCND corpus, GLoRIA outperforms geo-conditioned full fine-tuning, LoRA, and both dialect-specific and unified full fine-tuning, achieving state-of-the-art word error rates while updating under 10% of parameters. GLoRIA also generalizes well to unseen dialects, including in extrapolation scenarios, and enables interpretable adaptation patterns that can be visualized geospatially. These results show metadata-gated low-rank adaptation is an effective, interpretable, and efficient solution for dialectal ASR.

Wim Boes, Hugo Van hamme

In this technical report, the systems we submitted for subtask 1B of the DCASE 2021 challenge, regarding audiovisual scene classification, are described in detail. They are essentially multi-source transformers employing a combination of auditory and visual features to make predictions. These models are evaluated utilizing the macro-averaged multi-class cross-entropy and accuracy metrics. In terms of the macro-averaged multi-class cross-entropy, our best model achieved a score of 0.620 on the validation data. This is slightly better than the performance of the baseline system (0.658). With regard to the accuracy measure, our best model achieved a score of 77.1\% on the validation data, which is about the same as the performance obtained by the baseline system (77.0\%).

Wim Boes, Hugo Van hamme

Large-scale sound recognition data sets typically consist of acoustic recordings obtained from multimedia libraries. As a consequence, modalities other than audio can often be exploited to improve the outputs of models designed for associated tasks. Frequently, however, not all contents are available for all samples of such a collection: For example, the original material may have been removed from the source platform at some point, and therefore, non-auditory features can no longer be acquired. We demonstrate that a multi-encoder framework can be employed to deal with this issue by applying this method to attention-based deep learning systems, which are currently part of the state of the art in the domain of sound recognition. More specifically, we show that the proposed model extension can successfully be utilized to incorporate partially available visual information into the operational procedures of such networks, which normally only use auditory features during training and inference. Experimentally, we verify that the considered approach leads to improved predictions in a number of evaluation scenarios pertaining to audio tagging and sound event detection. Additionally, we scrutinize some properties and limitations of the presented technique.

Quentin Meeus, Marie-Francine Moens, Hugo Van hamme

While extensively explored in text-based tasks, Named Entity Recognition (NER) remains largely neglected in spoken language understanding. Existing resources are limited to a single, English-only dataset. This paper addresses this gap by introducing MSNER, a freely available, multilingual speech corpus annotated with named entities. It provides annotations to the VoxPopuli dataset in four languages (Dutch, French, German, and Spanish). We have also releasing an efficient annotation tool that leverages automatic pre-annotations for faster manual refinement. This results in 590 and 15 hours of silver-annotated speech for training and validation, alongside a 17-hour, manually-annotated evaluation set. We further provide an analysis comparing silver and gold annotations. Finally, we present baseline NER models to stimulate further research on this newly available dataset.

Henry Grafé, Hugo Van hamme

Automatic Phoneme Recognition (APR) systems are often trained using pseudo phoneme-level annotations generated from text through Grapheme-to-Phoneme (G2P) systems. These G2P systems frequently output multiple possible pronunciations per word, but the standard Connectionist Temporal Classification (CTC) loss cannot account for such ambiguity during training. In this work, we adapt Graph Temporal Classification (GTC) to the APR setting. GTC enables training from a graph of alternative phoneme sequences, allowing the model to consider multiple pronunciations per word as valid supervision. Our experiments on English and Dutch data sets show that incorporating multiple pronunciations per word into the training loss consistently improves phoneme error rates compared to a baseline trained with CTC. These results suggest that integrating pronunciation variation into the loss function is a promising strategy for training APR systems from noisy G2P-based supervision.

Pu Wang, Shinji Watanabe, Hugo Van hamme

Parameter-efficient fine-tuning (PEFT) has emerged as a scalable solution for adapting large foundation models. While low-rank adaptation (LoRA) is widely used in speech applications, its state-of-the-art variants, e.g., VeRA, DoRA, PiSSA, and SVFT, are developed mainly for language and vision tasks, with limited validation in speech. This work presents the first comprehensive integration and benchmarking of these PEFT methods within ESPnet. We further introduce structured SVD-guided (SSVD) fine-tuning, which selectively rotates input-associated right singular vectors while keeping output-associated vectors fixed to preserve semantic mappings. This design enables robust domain adaptation with minimal trainable parameters and improved efficiency. We evaluate all methods on domain-shifted speech recognition tasks, including child speech and dialectal variation, across model scales from 0.1B to 2B. All implementations are released in ESPnet to support reproducibility and future work.

Xinnian Zhao, Hugo Van Hamme

This study proposes a novel approach to using TV subtitles within a weakly supervised (WS) Automatic Speech Recognition (ASR) framework. Although TV subtitles are readily available, their imprecise alignment with corresponding audio limits their applicability as supervised targets for verbatim transcription. Rather than using subtitles as direct supervision signals, our method reimagines them as context-rich prompts. This design enables the model to handle discrepancies between spoken audio and subtitle text. Instead, generated pseudo transcripts become the primary targets, with subtitles acting as guiding cues for iterative refinement. To further enhance the process, we introduce a weighted attention mechanism that emphasizes relevant subtitle tokens during inference. Our experiments demonstrate significant improvements in transcription accuracy, highlighting the effectiveness of the proposed method in refining transcripts. These enhanced pseudo-labeled datasets provide high-quality foundational resources for training robust ASR systems.

Anyu Ying, Natarajan Balaji Shankar, Chyi-Jiunn Lin et al.

Despite advancements in ASR, child speech recognition remains challenging due to acoustic variability and limited annotated data. While fine-tuning adult ASR models on child speech is common, comparisons with flat-start training remain underexplored. We compare flat-start training across multiple datasets, SSL representations (WavLM, XEUS), and decoder architectures. Our results show that SSL representations are biased toward adult speech, with flat-start training on child speech mitigating these biases. We also analyze model scaling, finding consistent improvements up to 1B parameters, beyond which performance plateaus. Additionally, age-related ASR and speaker verification analysis highlights the limitations of proprietary models like Whisper, emphasizing the need for open-data models for reliable child speech research. All investigations are conducted using ESPnet, and our publicly available benchmark provides insights into training strategies for robust child speech processing.

Steven Vander Eeckt, Hugo Van hamme

Catastrophic forgetting remains a major challenge when neural networks learn tasks sequentially. Elastic Weight Consolidation (EWC) attempts to address this problem by introducing a Bayesian-inspired regularization loss to preserve knowledge of previously learned tasks. However, EWC relies on a Laplace approximation where the Hessian is simplified to the diagonal of the Fisher information matrix, assuming uncorrelated model parameters. This overly simplistic assumption often leads to poor Hessian estimates, limiting its effectiveness. To overcome this limitation, we introduce Continual Learning with Sampled Quasi-Newton (CSQN), which leverages Quasi-Newton methods to compute more accurate Hessian approximations. CSQN captures parameter interactions beyond the diagonal without requiring architecture-specific modifications, making it applicable across diverse tasks and architectures. Experimental results across four benchmarks demonstrate that CSQN consistently outperforms EWC and other state-of-the-art baselines, including rehearsal-based methods. CSQN reduces EWC's forgetting by 50 percent and improves its performance by 8 percent on average. Notably, CSQN achieves superior results on three out of four benchmarks, including the most challenging scenarios, highlighting its potential as a robust solution for continual learning.

Steven Vander Eeckt, Hugo Van hamme

Adapting Automatic Speech Recognition (ASR) models to new domains results in a deterioration of performance on the original domain(s), a phenomenon called Catastrophic Forgetting (CF). Even monolingual ASR models cannot be extended to new accents, dialects, topics, etc. without suffering from CF, making them unable to be continually enhanced without storing all past data. Fortunately, Continual Learning (CL) methods, which aim to enable continual adaptation while overcoming CF, can be used. In this paper, we implement an extensive number of CL methods for End-to-End ASR and test and compare their ability to extend a monolingual Hybrid CTC-Transformer model across four new tasks. We find that the best performing CL method closes the gap between the fine-tuned model (lower bound) and the model trained jointly on all tasks (upper bound) by more than 40%, while requiring access to only 0.6% of the original data.

Wei Liu, Meng Sun, Xiongwei Zhang et al.

The choice of an optimal time-frequency resolution is usually a difficult but important step in tasks involving speech signal classification, e.g., speech anti-spoofing. The variations of the performance with different choices of timefrequency resolutions can be as large as those with different model architectures, which makes it difficult to judge what the improvement actually comes from when a new network architecture is invented and introduced as the classifier. In this paper, we propose a multi-resolution front-end for feature extraction in an end-to-end classification framework. Optimal weighted combinations of multiple time-frequency resolutions will be learned automatically given the objective of a classification task. Features extracted with different time-frequency resolutions are weighted and concatenated as inputs to the successive networks, where the weights are predicted by a learnable neural network inspired by the weighting block in squeeze-and-excitation networks (SENet). Furthermore, the refinement of the chosen timefrequency resolutions is investigated by pruning the ones with relatively low importance, which reduces the complexity and size of the model. The proposed method is evaluated on the tasks of speech anti-spoofing in ASVSpoof 2019 and its superiority has been justified by comparing with similar baselines.

Jakob Poncelet, Hugo Van hamme

Recent research in speech processing exhibits a growing interest in unsupervised and self-supervised representation learning from unlabelled data to alleviate the need for large amounts of annotated data. We investigate several popular pre-training methods and apply them to Flemish Dutch. We compare off-the-shelf English pre-trained models to models trained on an increasing amount of Flemish data. We find that the most important factors for positive transfer to downstream speech recognition tasks include a substantial amount of data and a matching pre-training domain. Ideally, we also finetune on an annotated subset in the target language. All pre-trained models improve linear phone separability in Flemish, but not all methods improve Automatic Speech Recognition. We experience superior performance with wav2vec 2.0 and we obtain a 30% WER improvement by finetuning the multilingually pre-trained XLSR-53 model on Flemish Dutch, after integration into an HMM-DNN acoustic model.

Mohammad Jalilpour Monesi, Bernd Accou, Tom Francart et al.

Decoding the speech signal that a person is listening to from the human brain via electroencephalography (EEG) can help us understand how our auditory system works. Linear models have been used to reconstruct the EEG from speech or vice versa. Recently, Artificial Neural Networks (ANNs) such as Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) based architectures have outperformed linear models in modeling the relation between EEG and speech. Before attempting to use these models in real-world applications such as hearing tests or (second) language comprehension assessment we need to know what level of speech information is being utilized by these models. In this study, we aim to analyze the performance of an LSTM-based model using different levels of speech features. The task of the model is to determine which of two given speech segments is matched with the recorded EEG. We used low- and high-level speech features including: envelope, mel spectrogram, voice activity, phoneme identity, and word embedding. Our results suggest that the model exploits information about silences, intensity, and broad phonetic classes from the EEG. Furthermore, the mel spectrogram, which contains all this information, yields the highest accuracy (84%) among all the features.

Wim Boes, Robbe Van Rompaey, Lyan Verwimp et al.

We inspect the long-term learning ability of Long Short-Term Memory language models (LSTM LMs) by evaluating a contextual extension based on the Continuous Bag-of-Words (CBOW) model for both sentence- and discourse-level LSTM LMs and by analyzing its performance. We evaluate on text and speech. Sentence-level models using the long-term contextual module perform comparably to vanilla discourse-level LSTM LMs. On the other hand, the extension does not provide gains for discourse-level models. These findings indicate that discourse-level LSTM LMs already rely on contextual information to perform long-term learning.

Wim Boes, Hugo Van hamme

We study the merit of transfer learning for two sound recognition problems, i.e., audio tagging and sound event detection. Employing feature fusion, we adapt a baseline system utilizing only spectral acoustic inputs to also make use of pretrained auditory and visual features, extracted from networks built for different tasks and trained with external data. We perform experiments with these modified models on an audiovisual multi-label data set, of which the training partition contains a large number of unlabeled samples and a smaller amount of clips with weak annotations, indicating the clip-level presence of 10 sound categories without specifying the temporal boundaries of the active auditory events. For clip-based audio tagging, this transfer learning method grants marked improvements. Addition of the visual modality on top of audio also proves to be advantageous in this context. When it comes to generating transcriptions of audio recordings, the benefit of pretrained features depends on the requested temporal resolution: for coarse-grained sound event detection, their utility remains notable. But when more fine-grained predictions are required, performance gains are strongly reduced due to a mismatch between the problem at hand and the goals of the models from which the pretrained vectors were obtained.

Bernd Accou, Mohammad Jalilpour Monesi, Hugo Van hamme et al.

Objective: Currently, only behavioral speech understanding tests are available, which require active participation of the person being tested. As this is infeasible for certain populations, an objective measure of speech intelligibility is required. Recently, brain imaging data has been used to establish a relationship between stimulus and brain response. Linear models have been successfully linked to speech intelligibility but require per-subject training. We present a deep-learning-based model incorporating dilated convolutions that operates in a match/mismatch paradigm. The accuracy of the model's match/mismatch predictions can be used as a proxy for speech intelligibility without subject-specific (re)training. Approach: We evaluated the performance of the model as a function of input segment length, EEG frequency band and receptive field size while comparing it to multiple baseline models. Next, we evaluated performance on held-out data and finetuning. Finally, we established a link between the accuracy of our model and the state-of-the-art behavioral MATRIX test. Main results: The dilated convolutional model significantly outperformed the baseline models for every input segment length, for all EEG frequency bands except the delta and theta band, and receptive field sizes between 250 and 500 ms. Additionally, finetuning significantly increased the accuracy on a held-out dataset. Finally, a significant correlation (r=0.59, p=0.0154) was found between the speech reception threshold estimated using the behavioral MATRIX test and our objective method. Significance: Our method is the first to predict the speech reception threshold from EEG for unseen subjects, contributing to objective measures of speech intelligibility.

Pu Wang, Hugo Van hamme

Designing a speech-to-intent (S2I) agent which maps the users' spoken commands to the agents' desired task actions can be challenging due to the diverse grammatical and lexical preference of different users. As a remedy, we discuss a user-taught S2I system in this paper. The user-taught system learns from scratch from the users' spoken input with action demonstration, which ensure it is fully matched to the users' way of formulating intents and their articulation habits. The main issue is the scarce training data due to the user effort involved. Existing state-of-art approaches in this setting are based on non-negative matrix factorization (NMF) and capsule networks. In this paper we combine the encoder of an end-to-end ASR system with the prior NMF/capsule network-based user-taught decoder, and investigate whether pre-training methodology can reduce training data requirements for the NMF and capsule network. Experimental results show the pre-trained ASR-NMF framework significantly outperforms other models, and also, we discuss limitations of pre-training with different types of command-and-control(C&C) applications.

Jeroen Zegers, Hugo Van hamme

Long short-term memory recurrent neural networks (LSTM-RNNs) are considered state-of-the art in many speech processing tasks. The recurrence in the network, in principle, allows any input to be remembered for an indefinite time, a feature very useful for sequential data like speech. However, very little is known about which information is actually stored in the LSTM and for how long. We address this problem by using a memory reset approach which allows us to evaluate network performance depending on the allowed memory time span. We apply this approach to the task of multi-speaker source separation, but it can be used for any task using RNNs. We find a strong performance effect of short-term (shorter than 100 milliseconds) linguistic processes. Only speaker characteristics are kept in the memory for longer than 400 milliseconds. Furthermore, we confirm that performance-wise it is sufficient to implement longer memory in deeper layers. Finally, in a bidirectional model, the backward models contributes slightly more to the separation performance than the forward model.

Pieter Appeltans, Jeroen Zegers, Hugo Van hamme

This paper examines the applicability in realistic scenarios of two deep learning based solutions to the overlapping speaker separation problem. Firstly, we present experiments that show that these methods are applicable for a broad range of languages. Further experimentation indicates limited performance loss for untrained languages, when these have common features with the trained language(s). Secondly, it investigates how the methods deal with realistic background noise and proposes some modifications to better cope with these disturbances. The deep learning methods that will be examined are deep clustering and deep attractor networks.

Jeroen Zegers, Hugo Van hamme

In recent years there have been many deep learning approaches towards the multi-speaker source separation problem. Most use Long Short-Term Memory - Recurrent Neural Networks (LSTM-RNN) or Convolutional Neural Networks (CNN) to model the sequential behavior of speech. In this paper we propose a novel network for source separation using an encoder-decoder CNN and LSTM in parallel. Hyper parameters have to be chosen for both parts of the network and they are potentially mutually dependent. Since hyper parameter grid search has a high computational burden, random search is often preferred. However, when sampling a new point in the hyper parameter space, it can potentially be very close to a previously evaluated point and thus give little additional information. Furthermore, random sampling is as likely to sample in a promising area as in an hyper space area dominated with poor performing models. Therefore, we use a Bayesian hyper parameter optimization technique and find that the parallel CNN-LSTM outperforms the LSTM-only and CNN-only model.

Wim Boes, Hugo Van hamme

We tackle the task of environmental event classification by drawing inspiration from the transformer neural network architecture used in machine translation. We modify this attention-based feedforward structure in such a way that allows the resulting model to use audio as well as video to compute sound event predictions. We perform extensive experiments with these adapted transformers on an audiovisual data set, obtained by appending relevant visual information to an existing large-scale weakly labeled audio collection. The employed multi-label data contains clip-level annotation indicating the presence or absence of 17 classes of environmental sounds, and does not include temporal information. We show that the proposed modified transformers strongly improve upon previously introduced models and in fact achieve state-of-the-art results. We also make a compelling case for devoting more attention to research in multimodal audiovisual classification by proving the usefulness of visual information for the task at hand,namely audio event recognition. In addition, we visualize internal attention patterns of the audiovisual transformers and in doing so demonstrate their potential for performing multimodal synchronization.

Janneke van de Loo, Jort F. Gemmeke, Guy De Pauw et al.

We present a framework for the induction of semantic frames from utterances in the context of an adaptive command-and-control interface. The system is trained on an individual user's utterances and the corresponding semantic frames representing controls. During training, no prior information on the alignment between utterance segments and frame slots and values is available. In addition, semantic frames in the training data can contain information that is not expressed in the utterances. To tackle this weakly supervised classification task, we propose a framework based on Hidden Markov Models (HMMs). Structural modifications, resulting in a hierarchical HMM, and an extension called expression sharing are introduced to minimize the amount of training time and effort required for the user. The dataset used for the present study is PATCOR, which contains commands uttered in the context of a vocally guided card game, Patience. Experiments were carried out on orthographic and phonetic transcriptions of commands, segmented on different levels of n-gram granularity. The experimental results show positive effects of all the studied system extensions, with some effect differences between the different input representations. Moreover, evaluation experiments on held-out data with the optimal system configuration show that the extended system is able to achieve high accuracies with relatively small amounts of training data.

Lyan Verwimp, Joris Pelemans, Hugo Van hamme et al.

Neural cache language models (LMs) extend the idea of regular cache language models by making the cache probability dependent on the similarity between the current context and the context of the words in the cache. We make an extensive comparison of 'regular' cache models with neural cache models, both in terms of perplexity and WER after rescoring first-pass ASR results. Furthermore, we propose two extensions to this neural cache model that make use of the content value/information weight of the word: firstly, combining the cache probability and LM probability with an information-weighted interpolation and secondly, selectively adding only content words to the cache. We obtain a 29.9%/32.1% (validation/test set) relative improvement in perplexity with respect to a baseline LSTM LM on the WikiText-2 dataset, outperforming previous work on neural cache LMs. Additionally, we observe significant WER reductions with respect to the baseline model on the WSJ ASR task.

Jeroen Zegers, Hugo Van hamme

With deep learning approaches becoming state-of-the-art in many speech (as well as non-speech) related machine learning tasks, efforts are being taken to delve into the neural networks which are often considered as a black box. In this paper it is analyzed how recurrent neural network (RNNs) cope with temporal dependencies by determining the relevant memory time span in a long short-term memory (LSTM) cell. This is done by leaking the state variable with a controlled lifetime and evaluating the task performance. This technique can be used for any task to estimate the time span the LSTM exploits in that specific scenario. The focus in this paper is on the task of separating speakers from overlapping speech. We discern two effects: A long term effect, probably due to speaker characterization and a short term effect, probably exploiting phone-size formant tracks.

Jeroen Zegers, Hugo Van hamme

Research in deep learning for multi-speaker source separation has received a boost in the last years. However, most studies are restricted to mixtures of a specific number of speakers, called a specific scenario. While some works included experiments for different scenarios, research towards combining data of different scenarios or creating a single model for multiple scenarios have been very rare. In this work it is shown that data of a specific scenario is relevant for solving another scenario. Furthermore, it is concluded that a single model, trained on different scenarios is capable of matching performance of scenario specific models.

Lyan Verwimp, Hugo Van hamme, Vincent Renkens et al.

We present a framework for analyzing what the state in RNNs remembers from its input embeddings. Our approach is inspired by backpropagation, in the sense that we compute the gradients of the states with respect to the input embeddings. The gradient matrix is decomposed with Singular Value Decomposition to analyze which directions in the embedding space are best transferred to the hidden state space, characterized by the largest singular values. We apply our approach to LSTM language models and investigate to what extent and for how long certain classes of words are remembered on average for a certain corpus. Additionally, the extent to which a specific property or relationship is remembered by the RNN can be tracked by comparing a vector characterizing that property with the direction(s) in embedding space that are best preserved in hidden state space.

Vincent Renkens, Hugo Van hamme

Designing a spoken language understanding system for command-and-control applications can be challenging because of a wide variety of domains and users or because of a lack of training data. In this paper we discuss a system that learns from scratch from user demonstrations. This method has the advantage that the same system can be used for many domains and users without modifications and that no training data is required prior to deployment. The user is required to train the system, so for a user friendly experience it is crucial to minimize the required amount of data. In this paper we investigate whether a capsule network can make efficient use of the limited amount of available training data. We compare the proposed model to an approach based on Non-negative Matrix Factorisation which is the state-of-the-art in this setting and another deep learning approach that was recently introduced for end-to-end spoken language understanding. We show that the proposed model outperforms the baseline models for three command-and-control applications: controlling a small robot, a vocally guided card game and a home automation task.

Lyan Verwimp, Joris Pelemans, Marieke Lycke et al.

In Flanders, all TV shows are subtitled. However, the process of subtitling is a very time-consuming one and can be sped up by providing the output of a speech recognizer run on the audio of the TV show, prior to the subtitling. Naturally, this speech recognition will perform much better if the employed language model is adapted to the register and the topic of the program. We present several language models trained on subtitles of television shows provided by the Flemish public-service broadcaster VRT. This data was gathered in the context of the project STON which has as purpose to facilitate the process of subtitling TV shows. One model is trained on all available data (46M word tokens), but we also trained models on a specific type of TV show or domain/topic. Language models of spoken language are quite rare due to the lack of training data. The size of this corpus is relatively large for a corpus of spoken language (compare with e.g. CGN which has 9M words), but still rather small for a language model. Thus, in practice it is advised to interpolate these models with a large background language model trained on written language. The models can be freely downloaded on http://www.esat.kuleuven.be/psi/spraak/downloads/.

Jeroen Zegers, Hugo Van hamme

Lately there have been novel developments in deep learning towards solving the cocktail party problem. Initial results are very promising and allow for more research in the domain. One technique that has not yet been explored in the neural network approach to this task is speaker adaptation. Intuitively, information on the speakers that we are trying to separate seems fundamentally important for the speaker separation task. However, retrieving this speaker information is challenging since the speaker identities are not known a priori and multiple speakers are simultaneously active. There is thus some sort of chicken and egg problem. To tackle this, source signals and i-vectors are estimated alternately. We show that blind multi-speaker adaptation improves the results of the network and that (in our case) the network is not capable of adequately retrieving this useful speaker information itself.

Lyan Verwimp, Joris Pelemans, Hugo Van hamme et al.

We present a Character-Word Long Short-Term Memory Language Model which both reduces the perplexity with respect to a baseline word-level language model and reduces the number of parameters of the model. Character information can reveal structural (dis)similarities between words and can even be used when a word is out-of-vocabulary, thus improving the modeling of infrequent and unknown words. By concatenating word and character embeddings, we achieve up to 2.77% relative improvement on English compared to a baseline model with a similar amount of parameters and 4.57% on Dutch. Moreover, we also outperform baseline word-level models with a larger number of parameters.

Jeroen Zegers, Hugo Van hamme

Non-negative Matrix Factorization (NMF) has already been applied to learn speaker characterizations from single or non-simultaneous speech for speaker recognition applications. It is also known for its good performance in (blind) source separation for simultaneous speech. This paper explains how NMF can be used to jointly solve the two problems in a multichannel speaker recognizer for simultaneous speech. It is shown how state-of-the-art multichannel NMF for blind source separation can be easily extended to incorporate speaker recognition. Experiments on the CHiME corpus show that this method outperforms the sequential approach of first applying source separation, followed by speaker recognition that uses state-of-the-art i-vector techniques.

Hugo Van hamme

Non-negative matrix factorization (NMF) has become a popular machine learning approach to many problems in text mining, speech and image processing, bio-informatics and seismic data analysis to name a few. In NMF, a matrix of non-negative data is approximated by the low-rank product of two matrices with non-negative entries. In this paper, the approximation quality is measured by the Kullback-Leibler divergence between the data and its low-rank reconstruction. The existence of the simple multiplicative update (MU) algorithm for computing the matrix factors has contributed to the success of NMF. Despite the availability of algorithms showing faster convergence, MU remains popular due to its simplicity. In this paper, a diagonalized Newton algorithm (DNA) is proposed showing faster convergence while the implementation remains simple and suitable for high-rank problems. The DNA algorithm is applied to various publicly available data sets, showing a substantial speed-up on modern hardware.