Okko Räsänen

Marvin Lavechin, Yaya Sy, Hadrien Titeux et al.

Self-supervised techniques for learning speech representations have been shown to develop linguistic competence from exposure to speech without the need for human labels. In order to fully realize the potential of these approaches and further our understanding of how infants learn language, simulations must closely emulate real-life situations by training on developmentally plausible corpora and benchmarking against appropriate test sets. To this end, we propose a language-acquisition-friendly benchmark to probe spoken language models at the lexical and syntactic levels, both of which are compatible with the vocabulary typical of children's language experiences. This paper introduces the benchmark and summarizes a range of experiments showing its usefulness. In addition, we highlight two exciting challenges that need to be addressed for further progress: bridging the gap between text and speech and between clean speech and in-the-wild speech.

Khazar Khorrami, María Andrea Cruz Blandón, Tuomas Virtanen et al.

Speech representation learning with self-supervised algorithms has resulted in notable performance boosts in many downstream tasks. Recent work combined self-supervised learning (SSL) and visually grounded speech (VGS) processing mechanisms for representation learning. The joint training with SSL and VGS mechanisms provides the opportunity to utilize both unlabeled speech and speech-related visual information based on data availability. This has shown to enhance the quality of learned representations, especially at encoding semantic- and lexical-level knowledge. In this work, we further study the joint optimization of wav2vec 2.0-based SSL and transformer-based VGS as a multi-task learning system. We explore a set of training scenarios to understand how speech representations are shared or transferred between the two tasks, and what is the optimal training strategy for cross-modal semantic retrieval and phoneme discrimination performance. As a result, we find that sequential training with wav2vec 2.0 first and VGS next provides higher performance on audio-visual retrieval compared to simultaneous optimization of both learning mechanisms. However, the parallel SSL-VGS training reduces the effects of catastrophic forgetting when switching between optimization criteria. Moreover, the results suggest that phonemic representations learned through the VGS mechanism may generalize better across datasets compared to those learned with SSL.

Einari Vaaras, Manu Airaksinen, Okko Räsänen

When domain experts are needed to perform data annotation for complex machine-learning tasks, reducing annotation effort is crucial in order to cut down time and expenses. For cases when there are no annotations available, one approach is to utilize the structure of the feature space for clustering-based active learning (AL) methods. However, these methods are heavily dependent on how the samples are organized in the feature space and what distance metric is used. Unsupervised methods such as contrastive predictive coding (CPC) can potentially be used to learn organized feature spaces, but these methods typically create high-dimensional features which might be challenging for estimating data density. In this paper, we combine CPC and multiple dimensionality reduction methods in search of functioning practices for clustering-based AL. Our experiments for simulating speech emotion recognition system deployment show that both the local and global topology of the feature space can be successfully used for AL, and that CPC can be used to improve clustering-based AL performance over traditional signal features. Additionally, we observe that compressing data dimensionality does not harm AL performance substantially, and that 2-D feature representations achieved similar AL performance as higher-dimensional representations when the number of annotations is not very low.

Kalle Lahtinen, Einari Vaaras, Liisa Mustanoja et al.

Study of affect in speech requires suitable data, as emotional expression and perception vary across languages. Until now, no corpus has existed for natural expression of affect in spontaneous Finnish, existing data being acted or from a very specific communicative setting. This paper presents the first such corpus, created by annotating 12,000 utterances for emotional arousal and valence, sampled from three large-scale Finnish speech corpora. To ensure diverse affective expression, sample selection was conducted with an affect mining approach combining acoustic, cross-linguistic speech emotion, and text sentiment features. We compare this method to random sampling in terms of annotation diversity, and conduct post-hoc analyses to identify sampling choices that would have maximized the diversity. As an outcome, the work introduces a spontaneous Finnish affective speech corpus and informs sampling strategies for affective speech corpus creation in other languages or domains.

Einari Vaaras, Manu Airaksinen, Okko Räsänen

Self-supervised learning (SSL) is a data-driven learning approach that utilizes the innate structure of the data to guide the learning process. In contrast to supervised learning, which depends on external labels, SSL utilizes the inherent characteristics of the data to produce its own supervisory signal. However, one frequent issue with SSL methods is representation collapse, where the model outputs a constant input-invariant feature representation. This issue hinders the potential application of SSL methods to new data modalities, as trying to avoid representation collapse wastes researchers' time and effort. This paper introduces a novel SSL algorithm for time-series data called Prediction of Functionals from Masked Latents (PFML). Instead of predicting masked input signals or their latent representations directly, PFML operates by predicting statistical functionals of the input signal corresponding to masked embeddings, given a sequence of unmasked embeddings. The algorithm is designed to avoid representation collapse, rendering it straightforwardly applicable to different time-series data domains, such as novel sensor modalities in clinical data. We demonstrate the effectiveness of PFML through complex, real-life classification tasks across three different data modalities: infant posture and movement classification from multi-sensor inertial measurement unit data, emotion recognition from speech data, and sleep stage classification from EEG data. The results show that PFML is superior to a conceptually similar SSL method and a contrastive learning-based SSL method. Additionally, PFML is on par with the current state-of-the-art SSL method, while also being conceptually simpler and without suffering from representation collapse.

Daniil Kocharov, Okko Räsänen

Longform audio recordings obtained with microphones worn by children-also known as child-centered daylong recordings-have become a standard method for studying children's language experiences and their impact on subsequent language development. Transcripts of longform speech audio would enable rich analyses at various linguistic levels, yet the massive scale of typical longform corpora prohibits comprehensive manual annotation. At the same time, automatic speech recognition (ASR)-based transcription faces significant challenges due to the noisy, unconstrained nature of real-world audio, and no existing study has successfully applied ASR to transcribe such data. However, previous attempts have assumed that ASR must process each longform recording in its entirety. In this work, we present an approach to automatically detect those utterances in longform audio that can be reliably transcribed with modern ASR systems, allowing automatic and relatively accurate transcription of a notable proportion of all speech in typical longform data. We validate the approach on four English longform audio corpora, showing that it achieves a median word error rate (WER) of 0% and a mean WER of 18% when transcribing 13% of the total speech in the dataset. In contrast, transcribing all speech without any filtering yields a median WER of 52% and a mean WER of 51%. We also compare word log-frequencies derived from the automatic transcripts with those from manual annotations and show that the frequencies correlate at r = 0.92 (Pearson) for all transcribed words and r = 0.98 for words that appear at least five times in the automatic transcripts. Overall, the work provides a concrete step toward increasingly detailed automated linguistic analyses of child-centered longform audio.

Khazar Khorrami, Okko Räsänen

Infants gradually learn to parse continuous speech into words and connect names with objects, yet the mechanisms behind development of early word perception skills remain unknown. We studied the extent to which early words can be acquired through statistical learning from regularities in audiovisual sensory input. We simulated word learning in infants up to 12 months of age in a realistic setting, using a model that solely learns from statistical regularities in unannotated raw speech and pixel-level visual input. Crucially, the quantity of object naming events was carefully designed to match that accessible to infants of comparable ages. Results show that the model effectively learns to recognize words and associate them with corresponding visual objects, with a vocabulary growth rate comparable to that observed in infants. The findings support the viability of general statistical learning for early word perception, demonstrating how learning can operate without assuming any prior linguistic capabilities.

Okko Räsänen, Daniil Kocharov

Child-directed speech (CDS) is a particular type of speech that adults use when addressing young children. Its properties also change as a function of extralinguistic factors, such as age of the child being addressed. Access to large amounts of representative and varied CDS would be useful for child language research, as this would enable controlled computational modeling experiments of infant language acquisition with realistic input in terms of quality and quantity. In this study, we describe an approach to model age-dependent linguistic properties of CDS using a language model (LM) trained on CDS transcripts and ages of the recipient children, as obtained from North American English corpora of the CHILDES database. The created LM can then be used to stochastically generate synthetic CDS transcripts in an age-appropriate manner, thereby scaling beyond the original datasets in size. We compare characteristics of the generated CDS against the real speech addressed at children of different ages, showing that the LM manages to capture age-dependent changes in CDS, except for a slight difference in the effective vocabulary size. As a side product, we also provide a systematic characterization of age-dependent linguistic properties of CDS in CHILDES, illustrating how all measured aspects of the CDS change with children's age.

Puyuan Peng, Shang-Wen Li, Okko Räsänen et al.

In this paper, we show that representations capturing syllabic units emerge when training a self-supervised speech model with a visually-grounded training objective. We demonstrate that a nearly identical model architecture (HuBERT) trained with a masked language modeling loss does not exhibit this same ability, suggesting that the visual grounding objective is responsible for the emergence of this phenomenon. We propose the use of a minimum cut algorithm to automatically predict syllable boundaries in speech, followed by a 2-stage clustering method to group identical syllables together. We show that our model not only outperforms a state-of-the-art syllabic segmentation method on the language it was trained on (English), but also generalizes in a zero-shot fashion to Estonian. Finally, we show that the same model is capable of zero-shot generalization for a word segmentation task on 4 other languages from the Zerospeech Challenge, in some cases beating the previous state-of-the-art.

Einari Vaaras, Manu Airaksinen, Sampsa Vanhatalo et al.

The recently-developed infant wearable MAIJU provides a means to automatically evaluate infants' motor performance in an objective and scalable manner in out-of-hospital settings. This information could be used for developmental research and to support clinical decision-making, such as detection of developmental problems and guiding of their therapeutic interventions. MAIJU-based analyses rely fully on the classification of infant's posture and movement; it is hence essential to study ways to increase the accuracy of such classifications, aiming to increase the reliability and robustness of the automated analysis. Here, we investigated how self-supervised pre-training improves performance of the classifiers used for analyzing MAIJU recordings, and we studied whether performance of the classifier models is affected by context-selective quality-screening of pre-training data to exclude periods of little infant movement or with missing sensors. Our experiments show that i) pre-training the classifier with unlabeled data leads to a robust accuracy increase of subsequent classification models, and ii) selecting context-relevant pre-training data leads to substantial further improvements in the classifier performance.

María Andrea Cruz Blandón, Alejandrina Cristia, Okko Räsänen

Modelling of early language acquisition aims to understand how infants bootstrap their language skills. The modelling encompasses properties of the input data used for training the models, the cognitive hypotheses and their algorithmic implementations being tested, and the evaluation methodologies to compare models to human data. Recent developments have enabled the use of more naturalistic training data for computational models. This also motivates development of more naturalistic tests of model behaviour. A crucial step towards such an aim is to develop representative speech datasets consisting of speech heard by infants in their natural environments. However, a major drawback of such recordings is that they are typically noisy, and it is currently unclear how the sound quality could affect analyses and modelling experiments conducted on such data. In this paper, we explore this aspect for the case of infant-directed speech (IDS) and adult-directed speech (ADS) analysis. First, we manually and automatically annotated audio quality of utterances extracted from two corpora of child-centred long-form recordings (in English and French). We then compared acoustic features of IDS and ADS in an in-lab dataset and across different audio quality subsets of naturalistic data. Finally, we assessed how the audio quality and recording environment may change the conclusions of a modelling analysis using a recent self-supervised learning model. Our results show that the use of modest and high audio quality naturalistic speech data result in largely similar conclusions on IDS and ADS in terms of acoustic analyses and modelling experiments. We also found that an automatic sound quality assessment tool can be used to screen out useful parts of long-form recordings for a closer analysis with comparable results to that of manual quality annotation.

Kevin Eloff, Okko Räsänen, Herman A. Engelbrecht et al.

Multi-agent reinforcement learning has been used as an effective means to study emergent communication between agents, yet little focus has been given to continuous acoustic communication. This would be more akin to human language acquisition; human infants acquire language in large part through continuous signalling with their caregivers. We therefore ask: Are we able to observe emergent language between agents with a continuous communication channel? Our goal is to provide a platform to begin bridging the gap between human and agent communication, allowing us to analyse continuous signals, how they emerge, their characteristics, and how they relate to human language acquisition. We propose a messaging environment where a Speaker agent needs to convey a set of attributes to a Listener over a noisy acoustic channel. Using DQN to train our agents, we show that: (1) unlike the discrete case, the acoustic Speaker learns redundancy to improve Listener coherency, (2) the acoustic Speaker develops more compositional communication protocols which implicitly compensates for transmission errors over a noisy channel, and (3) DQN has significant performance gains and increased compositionality when compared to previous methods optimised using REINFORCE.

Khazar Khorrami, Okko Räsänen

Decades of research has studied how language learning infants learn to discriminate speech sounds, segment words, and associate words with their meanings. While gradual development of such capabilities is unquestionable, the exact nature of these skills and the underlying mental representations yet remains unclear. In parallel, computational studies have shown that basic comprehension of speech can be achieved by statistical learning between speech and concurrent referentially ambiguous visual input. These models can operate without prior linguistic knowledge such as representations of linguistic units, and without learning mechanisms specifically targeted at such units. This has raised the question of to what extent knowledge of linguistic units, such as phone(me)s, syllables, and words, could actually emerge as latent representations supporting the translation between speech and representations in other modalities, and without the units being proximal learning targets for the learner. In this study, we formulate this idea as the so-called latent language hypothesis (LLH), connecting linguistic representation learning to general predictive processing within and across sensory modalities. We review the extent that the audiovisual aspect of LLH is supported by the existing computational studies. We then explore LLH further in extensive learning simulations with different neural network models for audiovisual cross-situational learning, and comparing learning from both synthetic and real speech data. We investigate whether the latent representations learned by the networks reflect phonetic, syllabic, or lexical structure of input speech by utilizing an array of complementary evaluation metrics related to linguistic selectivity and temporal characteristics of the representations. As a result, we find that representations associated...

Yuanyuan Liu, Nelly Penttilä, Tiina Ihalainen et al.

Imprecise vowel articulation can be observed in people with Parkinson's disease (PD). Acoustic features measuring vowel articulation have been demonstrated to be effective indicators of PD in its assessment. Standard clinical vowel articulation features of vowel working space area (VSA), vowel articulation index (VAI) and formants centralization ratio (FCR), are derived the first two formants of the three corner vowels /a/, /i/ and /u/. Conventionally, manual annotation of the corner vowels from speech data is required before measuring vowel articulation. This process is time-consuming. The present work aims to reduce human effort in clinical analysis of PD speech by proposing an automatic pipeline for vowel articulation assessment. The method is based on automatic corner vowel detection using a language universal phoneme recognizer, followed by statistical analysis of the formant data. The approach removes the restrictions of prior knowledge of speaking content and the language in question. Experimental results on a Finnish PD speech corpus demonstrate the efficacy and reliability of the proposed automatic method in deriving VAI, VSA, FCR and F2i/F2u (the second formant ratio for vowels /i/ and /u/). The automatically computed parameters are shown to be highly correlated with features computed with manual annotations of corner vowels. In addition, automatically and manually computed vowel articulation features have comparable correlations with experts' ratings on speech intelligibility, voice impairment and overall severity of communication disorder. Language-independence of the proposed approach is further validated on a Spanish PD database, PC-GITA, as well as on TORGO corpus of English dysarthric speech.

Afra Alishahi, Grzegorz Chrupała, Alejandrina Cristia et al.

We present the visually-grounded language modelling track that was introduced in the Zero-Resource Speech challenge, 2021 edition, 2nd round. We motivate the new track and discuss participation rules in detail. We also present the two baseline systems that were developed for this track.

Khazar Khorrami, Okko Räsänen

Systems that can find correspondences between multiple modalities, such as between speech and images, have great potential to solve different recognition and data analysis tasks in an unsupervised manner. This work studies multimodal learning in the context of visually grounded speech (VGS) models, and focuses on their recently demonstrated capability to extract spatiotemporal alignments between spoken words and the corresponding visual objects without ever been explicitly trained for object localization or word recognition. As the main contributions, we formalize the alignment problem in terms of an audiovisual alignment tensor that is based on earlier VGS work, introduce systematic metrics for evaluating model performance in aligning visual objects and spoken words, and propose a new VGS model variant for the alignment task utilizing cross-modal attention layer. We test our model and a previously proposed model in the alignment task using SPEECH-COCO captions coupled with MSCOCO images. We compare the alignment performance using our proposed evaluation metrics to the semantic retrieval task commonly used to evaluate VGS models. We show that cross-modal attention layer not only helps the model to achieve higher semantic cross-modal retrieval performance, but also leads to substantial improvements in the alignment performance between image object and spoken words.

Manu Airaksinen, Sampsa Vanhatalo, Okko Räsänen

Infant motility assessment using intelligent wearables is a promising new approach for assessment of infant neurophysiological development, and where efficient signal analysis plays a central role. This study investigates the use of different end-to-end neural network architectures for processing infant motility data from wearable sensors. We focus on the performance and computational burden of alternative sensor encoder and time-series modelling modules and their combinations. In addition, we explore the benefits of data augmentation methods in ideal and non-ideal recording conditions. The experiments are conducted using a data-set of multi-sensor movement recordings from 7-month-old infants, as captured by a recently proposed smart jumpsuit for infant motility assessment. Our results indicate that the choice of the encoder module has a major impact on classifier performance. For sensor encoders, the best performance was obtained with parallel 2-dimensional convolutions for intra-sensor channel fusion with shared weights for all sensors. The results also indicate that a relatively compact feature representation is obtainable for within-sensor feature extraction without a drastic loss to classifier performance. Comparison of time-series models revealed that feed-forward dilated convolutions with residual and skip connections outperformed all RNN-based models in performance, training time, and training stability. The experiments also indicate that data augmentation improves model robustness in simulated packet loss or sensor dropout scenarios. In particular, signal- and sensor-dropout-based augmentation strategies provided considerable boosts to performance without negatively affecting the baseline performance. Overall the results provide tangible suggestions on how to optimize end-to-end neural network training for multi-channel movement sensor data.

Einari Vaaras, Sari Ahlqvist-Björkroth, Konstantinos Drossos et al.

Researchers have recently started to study how the emotional speech heard by young infants can affect their developmental outcomes. As a part of this research, hundreds of hours of daylong recordings from preterm infants' audio environments were collected from two hospitals in Finland and Estonia in the context of so-called APPLE study. In order to analyze the emotional content of speech in such a massive dataset, an automatic speech emotion recognition (SER) system is required. However, there are no emotion labels or existing indomain SER systems to be used for this purpose. In this paper, we introduce this initially unannotated large-scale real-world audio dataset and describe the development of a functional SER system for the Finnish subset of the data. We explore the effectiveness of alternative state-of-the-art techniques to deploy a SER system to a new domain, comparing cross-corpus generalization, WGAN-based domain adaptation, and active learning in the task. As a result, we show that the best-performing models are able to achieve a classification performance of 73.4% unweighted average recall (UAR) and 73.2% UAR for a binary classification for valence and arousal, respectively. The results also show that active learning achieves the most consistent performance compared to the two alternatives.

Okko Räsänen, María Andrea Cruz Blandón

Unsupervised spoken term discovery (UTD) aims at finding recurring segments of speech from a corpus of acoustic speech data. One potential approach to this problem is to use dynamic time warping (DTW) to find well-aligning patterns from the speech data. However, automatic selection of initial candidate segments for the DTW-alignment and detection of "sufficiently good" alignments among those require some type of pre-defined criteria, often operationalized as threshold parameters for pair-wise distance metrics between signal representations. In the existing UTD systems, the optimal hyperparameters may differ across datasets, limiting their applicability to new corpora and truly low-resource scenarios. In this paper, we propose a novel probabilistic approach to DTW-based UTD named as PDTW. In PDTW, distributional characteristics of the processed corpus are utilized for adaptive evaluation of alignment quality, thereby enabling systematic discovery of pattern pairs that have similarity what would be expected by coincidence. We test PDTW on Zero Resource Speech Challenge 2017 datasets as a part of 2020 implementation of the challenge. The results show that the system performs consistently on all five tested languages using fixed hyperparameters, clearly outperforming the earlier DTW-based system in terms of coverage of the detected patterns.

María Andrea Cruz Blandón, Okko Räsänen

Neural network models using predictive coding are interesting from the viewpoint of computational modelling of human language acquisition, where the objective is to understand how linguistic units could be learned from speech without any labels. Even though several promising predictive coding -based learning algorithms have been proposed in the literature, it is currently unclear how well they generalise to different languages and training dataset sizes. In addition, despite that such models have shown to be effective phonemic feature learners, it is unclear whether minimisation of the predictive loss functions of these models also leads to optimal phoneme-like representations. The present study investigates the behaviour of two predictive coding models, Autoregressive Predictive Coding and Contrastive Predictive Coding, in a phoneme discrimination task (ABX task) for two languages with different dataset sizes. Our experiments show a strong correlation between the autoregressive loss and the phoneme discrimination scores with the two datasets. However, to our surprise, the CPC model shows rapid convergence already after one pass over the training data, and, on average, its representations outperform those of APC on both languages.

Manu Airaksinen, Okko Räsänen, Elina Ilén et al.

Infants' spontaneous and voluntary movements mirror developmental integrity of brain networks since they require coordinated activation of multiple sites in the central nervous system. Accordingly, early detection of infants with atypical motor development holds promise for recognizing those infants who are at risk for a wide range of neurodevelopmental disorders (e.g., cerebral palsy, autism spectrum disorders). Previously, novel wearable technology has shown promise for offering efficient, scalable and automated methods for movement assessment in adults. Here, we describe the development of an infant wearable, a multi-sensor smart jumpsuit that allows mobile accelerometer and gyroscope data collection during movements. Using this suit, we first recorded play sessions of 22 typically developing infants of approximately 7 months of age. These data were manually annotated for infant posture and movement based on video recordings of the sessions, and using a novel annotation scheme specifically designed to assess the overall movement pattern of infants in the given age group. A machine learning algorithm, based on deep convolutional neural networks (CNNs) was then trained for automatic detection of posture and movement classes using the data and annotations. Our experiments show that the setup can be used for quantitative tracking of infant movement activities with a human equivalent accuracy, i.e., it meets the human inter-rater agreement levels in infant posture and movement classification. We also quantify the ambiguity of human observers in analyzing infant movements, and propose a method for utilizing this uncertainty for performance improvements in training of the automated classifier. Comparison of different sensor configurations also shows that four-limb recording leads to the best performance in posture and movement classification.

Okko Räsänen, Khazar Khorrami

Earlier research has suggested that human infants might use statistical dependencies between speech and non-linguistic multimodal input to bootstrap their language learning before they know how to segment words from running speech. However, feasibility of this hypothesis in terms of real-world infant experiences has remained unclear. This paper presents a step towards a more realistic test of the multimodal bootstrapping hypothesis by describing a neural network model that can learn word segments and their meanings from referentially ambiguous acoustic input. The model is tested on recordings of real infant-caregiver interactions using utterance-level labels for concrete visual objects that were attended by the infant when caregiver spoke an utterance containing the name of the object, and using random visual labels for utterances during absence of attention. The results show that beginnings of lexical knowledge may indeed emerge from individually ambiguous learning scenarios. In addition, the hidden layers of the network show gradually increasing selectivity to phonetic categories as a function of layer depth, resembling models trained for phone recognition in a supervised manner.

Shreyas Seshadri, Okko Räsänen

Automatic syllable count estimation (SCE) is used in a variety of applications ranging from speaking rate estimation to detecting social activity from wearable microphones or developmental research concerned with quantifying speech heard by language-learning children in different environments. The majority of previously utilized SCE methods have relied on heuristic DSP methods, and only a small number of bi-directional long short-term memory (BLSTM) approaches have made use of modern machine learning approaches in the SCE task. This paper presents a novel end-to-end method called SylNet for automatic syllable counting from speech, built on the basis of a recent developments in neural network architectures. We describe how the entire model can be optimized directly to minimize SCE error on the training data without annotations aligned at the syllable level, and how it can be adapted to new languages using limited speech data with known syllable counts. Experiments on several different languages reveal that SylNet generalizes to languages beyond its training data and further improves with adaptation. It also outperforms several previously proposed methods for syllabification, including end-to-end BLSTMs.

Paul Michel, Okko Räsänen, Roland Thiollière et al.

Phonemic segmentation of speech is a critical step of speech recognition systems. We propose a novel unsupervised algorithm based on sequence prediction models such as Markov chains and recurrent neural network. Our approach consists in analyzing the error profile of a model trained to predict speech features frame-by-frame. Specifically, we try to learn the dynamics of speech in the MFCC space and hypothesize boundaries from local maxima in the prediction error. We evaluate our system on the TIMIT dataset, with improvements over similar methods.