Herman Kamper

Simon Malan, Danel Slabbert, Herman Kamper

Building a lexicon from discovered word-like units is a central goal in zero-resource speech processing. But do our evaluations provide a trustworthy indication of lexicon quality? A common metric, normalized edit distance, averages the phoneme edit distances between discovered units in each cluster. We show that this metric has an inherent bias toward the quality of large clusters, inhibiting fair evaluation. Moreover, it ignores how well true classes are distributed across clusters. Based on established theory in clustering literature, we propose two metrics that address these shortcomings: a modified metric that weighs cluster size when assessing within-cluster consistency, and an inverse metric that assesses how true words are spread across clusters. Through experiments on synthetic and real-world lexicons, we demonstrate that combined, these metrics are: (1) more closely correlated with how similar a lexicon is to the ground-truth distribution, and (2) more robust to biases that skew lexicon evaluations.

Matthew Baas, Herman Kamper

We propose AudioStyleGAN (ASGAN), a new generative adversarial network (GAN) for unconditional speech synthesis. As in the StyleGAN family of image synthesis models, ASGAN maps sampled noise to a disentangled latent vector which is then mapped to a sequence of audio features so that signal aliasing is suppressed at every layer. To successfully train ASGAN, we introduce a number of new techniques, including a modification to adaptive discriminator augmentation to probabilistically skip discriminator updates. ASGAN achieves state-of-the-art results in unconditional speech synthesis on the Google Speech Commands dataset. It is also substantially faster than the top-performing diffusion models. Through a design that encourages disentanglement, ASGAN is able to perform voice conversion and speech editing without being explicitly trained to do so. ASGAN demonstrates that GANs are still highly competitive with diffusion models. Code, models, samples: https://github.com/RF5/simple-asgan/.

Benjamin van Niekerk, Marc-André Carbonneau, Herman Kamper

Voice conversion aims to transform source speech into a different target voice. However, typical voice conversion systems do not account for rhythm, which is an important factor in the perception of speaker identity. To bridge this gap, we introduce Urhythmic-an unsupervised method for rhythm conversion that does not require parallel data or text transcriptions. Using self-supervised representations, we first divide source audio into segments approximating sonorants, obstruents, and silences. Then we model rhythm by estimating speaking rate or the duration distribution of each segment type. Finally, we match the target speaking rate or rhythm by time-stretching the speech segments. Experiments show that Urhythmic outperforms existing unsupervised methods in terms of quality and prosody. Code and checkpoints: https://github.com/bshall/urhythmic. Audio demo page: https://ubisoft-laforge.github.io/speech/urhythmic.

Matthew Baas, Herman Kamper

Can we develop a model that can synthesize realistic speech directly from a latent space, without explicit conditioning? Despite several efforts over the last decade, previous adversarial and diffusion-based approaches still struggle to achieve this, even on small-vocabulary datasets. To address this, we propose AudioStyleGAN (ASGAN) -- a generative adversarial network for unconditional speech synthesis tailored to learn a disentangled latent space. Building upon the StyleGAN family of image synthesis models, ASGAN maps sampled noise to a disentangled latent vector which is then mapped to a sequence of audio features so that signal aliasing is suppressed at every layer. To successfully train ASGAN, we introduce a number of new techniques, including a modification to adaptive discriminator augmentation which probabilistically skips discriminator updates. We apply it on the small-vocabulary Google Speech Commands digits dataset, where it achieves state-of-the-art results in unconditional speech synthesis. It is also substantially faster than existing top-performing diffusion models. We confirm that ASGAN's latent space is disentangled: we demonstrate how simple linear operations in the space can be used to perform several tasks unseen during training. Specifically, we perform evaluations in voice conversion, speech enhancement, speaker verification, and keyword classification. Our work indicates that GANs are still highly competitive in the unconditional speech synthesis landscape, and that disentangled latent spaces can be used to aid generalization to unseen tasks. Code, models, samples: https://github.com/RF5/simple-asgan/

Matthew Baas, Kevin Eloff, Herman Kamper

Diffusion models have shown exceptional scaling properties in the image synthesis domain, and initial attempts have shown similar benefits for applying diffusion to unconditional text synthesis. Denoising diffusion models attempt to iteratively refine a sampled noise signal until it resembles a coherent signal (such as an image or written sentence). In this work we aim to see whether the benefits of diffusion models can also be realized for speech recognition. To this end, we propose a new way to perform speech recognition using a diffusion model conditioned on pretrained speech features. Specifically, we propose TransFusion: a transcribing diffusion model which iteratively denoises a random character sequence into coherent text corresponding to the transcript of a conditioning utterance. We demonstrate comparable performance to existing high-performing contrastive models on the LibriSpeech speech recognition benchmark. To the best of our knowledge, we are the first to apply denoising diffusion to speech recognition. We also propose new techniques for effectively sampling and decoding multinomial diffusion models. These are required because traditional methods of sampling from acoustic models are not possible with our new discrete diffusion approach. Code and trained models are available: https://github.com/RF5/transfusion-asr

Kayode Olaleye, Dan Oneata, Herman Kamper

Visually grounded speech (VGS) models are trained on images paired with unlabelled spoken captions. Such models could be used to build speech systems in settings where it is impossible to get labelled data, e.g. for documenting unwritten languages. However, most VGS studies are in English or other high-resource languages. This paper attempts to address this shortcoming. We collect and release a new single-speaker dataset of audio captions for 6k Flickr images in Yorùbá -- a real low-resource language spoken in Nigeria. We train an attention-based VGS model where images are automatically tagged with English visual labels and paired with Yorùbá utterances. This enables cross-lingual keyword localisation: a written English query is detected and located in Yorùbá speech. To quantify the effect of the smaller dataset, we compare to English systems trained on similar and more data. We hope that this new dataset will stimulate research in the use of VGS models for real low-resource languages.

Christiaan Jacobs, Nathanaël Carraz Rakotonirina, Everlyn Asiko Chimoto et al.

We consider hate speech detection through keyword spotting on radio broadcasts. One approach is to build an automatic speech recognition (ASR) system for the target low-resource language. We compare this to using acoustic word embedding (AWE) models that map speech segments to a space where matching words have similar vectors. We specifically use a multilingual AWE model trained on labelled data from well-resourced languages to spot keywords in data in the unseen target language. In contrast to ASR, the AWE approach only requires a few keyword exemplars. In controlled experiments on Wolof and Swahili where training and test data are from the same domain, an ASR model trained on just five minutes of data outperforms the AWE approach. But in an in-the-wild test on Swahili radio broadcasts with actual hate speech keywords, the AWE model (using one minute of template data) is more robust, giving similar performance to an ASR system trained on 30 hours of labelled data.

Matthew Baas, Herman Kamper

Voice conversion aims to convert source speech into a target voice using recordings of the target speaker as a reference. Newer models are producing increasingly realistic output. But what happens when models are fed with non-standard data, such as speech from a user with a speech impairment? We investigate how a recent voice conversion model performs on non-standard downstream voice conversion tasks. We use a simple but robust approach called k-nearest neighbors voice conversion (kNN-VC). We look at four non-standard applications: stuttered voice conversion, cross-lingual voice conversion, musical instrument conversion, and text-to-voice conversion. The latter involves converting to a target voice specified through a text description, e.g. "a young man with a high-pitched voice". Compared to an established baseline, we find that kNN-VC retains high performance in stuttered and cross-lingual voice conversion. Results are more mixed for the musical instrument and text-to-voice conversion tasks. E.g., kNN-VC works well on some instruments like drums but not on others. Nevertheless, this shows that voice conversion models - and kNN-VC in particular - are increasingly applicable in a range of non-standard downstream tasks. But there are still limitations when samples are very far from the training distribution. Code, samples, trained models: https://rf5.github.io/sacair2023-knnvc-demo/.

Leanne Nortje, Herman Kamper

Imagine being able to show a system a visual depiction of a keyword and finding spoken utterances that contain this keyword from a zero-resource speech corpus. We formalise this task and call it visually prompted keyword localisation (VPKL): given an image of a keyword, detect and predict where in an utterance the keyword occurs. To do VPKL, we propose a speech-vision model with a novel localising attention mechanism which we train with a new keyword sampling scheme. We show that these innovations give improvements in VPKL over an existing speech-vision model. We also compare to a visual bag-of-words (BoW) model where images are automatically tagged with visual labels and paired with unlabelled speech. Although this visual BoW can be queried directly with a written keyword (while our's takes image queries), our new model still outperforms the visual BoW in both detection and localisation, giving a 16% relative improvement in localisation F1.

Leanne Nortje, Dan Oneata, Herman Kamper

We propose a visually grounded speech model that learns new words and their visual depictions from just a few word-image example pairs. Given a set of test images and a spoken query, we ask the model which image depicts the query word. Previous work has simplified this few-shot learning problem by either using an artificial setting with digit word-image pairs or by using a large number of examples per class. Moreover, all previous studies were performed using English speech-image data. We propose an approach that can work on natural word-image pairs but with less examples, i.e. fewer shots, and then illustrate how this approach can be applied for multimodal few-shot learning in a real low-resource language, Yorùbá. Our approach involves using the given word-image example pairs to mine new unsupervised word-image training pairs from large collections of unlabelled speech and images. Additionally, we use a word-to-image attention mechanism to determine word-image similarity. With this new model, we achieve better performance with fewer shots than previous approaches on an existing English benchmark. Many of the model's mistakes are due to confusion between visual concepts co-occurring in similar contexts. The experiments on Yorùbá show the benefit of transferring knowledge from a multimodal model trained on a larger set of English speech-image data.

Simon Malan, Benjamin van Niekerk, Herman Kamper

We look at the long-standing problem of segmenting unlabeled speech into word-like segments and clustering these into a lexicon. Several previous methods use a scoring model coupled with dynamic programming to find an optimal segmentation. Here we propose a much simpler strategy: we predict word boundaries using the dissimilarity between adjacent self-supervised features, then we cluster the predicted segments to construct a lexicon. For a fair comparison, we update the older ES-KMeans dynamic programming method with better features and boundary constraints. On the five-language ZeroSpeech benchmarks, our simple approach gives similar state-of-the-art results compared to the new ES-KMeans+ method, while being almost five times faster. Project webpage: https://s-malan.github.io/prom-seg-clus.

Christiaan Jacobs, Herman Kamper

Acoustic word embeddings (AWEs) are fixed-dimensional vector representations of speech segments that encode phonetic content so that different realisations of the same word have similar embeddings. In this paper we explore semantic AWE modelling. These AWEs should not only capture phonetics but also the meaning of a word (similar to textual word embeddings). We consider the scenario where we only have untranscribed speech in a target language. We introduce a number of strategies leveraging a pre-trained multilingual AWE model -- a phonetic AWE model trained on labelled data from multiple languages excluding the target. Our best semantic AWE approach involves clustering word segments using the multilingual AWE model, deriving soft pseudo-word labels from the cluster centroids, and then training a Skipgram-like model on the soft vectors. In an intrinsic word similarity task measuring semantics, this multilingual transfer approach outperforms all previous semantic AWE methods. We also show -- for the first time -- that AWEs can be used for downstream semantic query-by-example search.

Werner van der Merwe, Herman Kamper, Johan du Preez

Latent Dirichlet allocation (LDA) is widely used for unsupervised topic modelling on sets of documents. No temporal information is used in the model. However, there is often a relationship between the corresponding topics of consecutive tokens. In this paper, we present an extension to LDA that uses a Markov chain to model temporal information. We use this new model for acoustic unit discovery from speech. As input tokens, the model takes a discretised encoding of speech from a vector quantised (VQ) neural network with 512 codes. The goal is then to map these 512 VQ codes to 50 phone-like units (topics) in order to more closely resemble true phones. In contrast to the base LDA, which only considers how VQ codes co-occur within utterances (documents), the Markov chain LDA additionally captures how consecutive codes follow one another. This extension leads to an increase in cluster quality and phone segmentation results compared to the base LDA. Compared to a recent vector quantised neural network approach that also learns 50 units, the extended LDA model performs better in phone segmentation but worse in mutual information.

Leanne Nortje, Dan Oneata, Gabriel Pirlogeanu et al.

Given an image query, visually prompted keyword localisation (VPKL) aims to find occurrences of the depicted word in a speech collection. This can be useful when transcriptions are not available for a low-resource language (e.g. if it is unwritten). Previous work showed that VPKL can be performed with a visually grounded speech model trained on paired images and unlabelled speech. But all experiments were done on English. Moreover, transcriptions were used to get positive and negative pairs for the contrastive loss. This paper introduces a few-shot learning scheme to mine pairs automatically without transcriptions. On English, this results in only a small drop in performance. We also - for the first time - consider VPKL on a real low-resource language, Yoruba. While scores are reasonable, here we see a bigger drop in performance compared to using ground truth pairs because the mining is less accurate in Yoruba.

Kyle Janse van Rensburg, Benjamin van Niekerk, Herman Kamper

How do speech models trained through self-supervised learning structure their representations? Previous studies have looked at how information is encoded in feature vectors across different layers. But few studies have considered whether speech characteristics are captured within individual dimensions of SSL features. In this paper we specifically look at speaker information using PCA on utterance-averaged representations. Using WavLM, we find that the principal dimension that explains most variance encodes pitch and associated characteristics like gender. Other individual principal dimensions correlate with intensity, noise levels, the second formant, and higher frequency characteristics. Finally, in synthesis experiments we show that most characteristics can be controlled by changing the corresponding dimensions. This provides a simple method to control characteristics of the output voice in synthesis applications.

Simon Malan, Benjamin van Niekerk, Herman Kamper

We investigate the problem of segmenting unlabeled speech into word-like units and clustering these to create a lexicon. Prior work can be categorized into two frameworks. Bottom-up methods first determine boundaries and then cluster the fixed segmented words into a lexicon. In contrast, top-down methods incorporate information from the clustered words to inform boundary selection. However, it is unclear whether top-down information is necessary to improve segmentation. To explore this, we look at two similar approaches that differ in whether top-down clustering informs boundary selection. Our simple bottom-up strategy predicts word boundaries using the dissimilarity between adjacent self-supervised features, then clusters the resulting segments to construct a lexicon. Our top-down system is an updated version of the ES-KMeans dynamic programming method that iteratively uses K-means to update its boundaries. On the five-language ZeroSpeech benchmarks, both approaches achieve comparable state-of-the-art results, with the bottom-up system being nearly five times faster. Through detailed analyses, we show that the top-down influence of ES-KMeans can be beneficial (depending on factors like the candidate boundaries), but in many cases the simple bottom-up method performs just as well. For both methods, we show that the clustering step is a limiting factor. Therefore, we recommend that future work focus on improved clustering techniques and learning more discriminative word-like representations. Project code repository: https://github.com/s-malan/prom-seg-clus.

Marc-André Carbonneau, Benjamin van Niekerk, Hugo Seuté et al.

Modeling voice identity is challenging due to its multifaceted nature. In generative speech systems, identity is often assessed using automatic speaker verification (ASV) embeddings, designed for discrimination rather than characterizing identity. This paper investigates which aspects of a voice are captured in such representations. We find that widely used ASV embeddings focus mainly on static features like timbre and pitch range, while neglecting dynamic elements such as rhythm. We also identify confounding factors that compromise speaker similarity measurements and suggest mitigation strategies. To address these gaps, we propose U3D, a metric that evaluates speakers' dynamic rhythm patterns. This work contributes to the ongoing challenge of assessing speaker identity consistency in the context of ever-better voice cloning systems. We publicly release our code.

Dan Oneata, Leanne Nortje, Yevgen Matusevych et al.

Mutual exclusivity (ME) is a strategy where a novel word is associated with a novel object rather than a familiar one, facilitating language learning in children. Recent work has found an ME bias in a visually grounded speech (VGS) model trained on English speech with paired images. But ME has also been studied in bilingual children, who may employ it less due to cross-lingual ambiguity. We explore this pattern computationally using bilingual VGS models trained on combinations of English, French, and Dutch. We find that bilingual models generally exhibit a weaker ME bias than monolingual models, though exceptions exist. Analyses show that the combined visual embeddings of bilingual models have a smaller variance for familiar data, partly explaining the increase in confusion between novel and familiar concepts. We also provide new insights into why the ME bias exists in VGS models in the first place. Code and data: https://github.com/danoneata/me-vgs

Benjamin van Niekerk, Marc-André Carbonneau, Julian Zaïdi et al.

The goal of voice conversion is to transform source speech into a target voice, keeping the content unchanged. In this paper, we focus on self-supervised representation learning for voice conversion. Specifically, we compare discrete and soft speech units as input features. We find that discrete representations effectively remove speaker information but discard some linguistic content - leading to mispronunciations. As a solution, we propose soft speech units. To learn soft units, we predict a distribution over discrete speech units. By modeling uncertainty, soft units capture more content information, improving the intelligibility and naturalness of converted speech. Samples available at https://ubisoft-laforge.github.io/speech/soft-vc/. Code available at https://github.com/bshall/soft-vc/.

Wilhelmina Nekoto, Vukosi Marivate, Tshinondiwa Matsila et al.

Research in NLP lacks geographic diversity, and the question of how NLP can be scaled to low-resourced languages has not yet been adequately solved. "Low-resourced"-ness is a complex problem going beyond data availability and reflects systemic problems in society. In this paper, we focus on the task of Machine Translation (MT), that plays a crucial role for information accessibility and communication worldwide. Despite immense improvements in MT over the past decade, MT is centered around a few high-resourced languages. As MT researchers cannot solve the problem of low-resourcedness alone, we propose participatory research as a means to involve all necessary agents required in the MT development process. We demonstrate the feasibility and scalability of participatory research with a case study on MT for African languages. Its implementation leads to a collection of novel translation datasets, MT benchmarks for over 30 languages, with human evaluations for a third of them, and enables participants without formal training to make a unique scientific contribution. Benchmarks, models, data, code, and evaluation results are released under https://github.com/masakhane-io/masakhane-mt.

Iroro Orife, Julia Kreutzer, Blessing Sibanda et al.

Africa has over 2000 languages. Despite this, African languages account for a small portion of available resources and publications in Natural Language Processing (NLP). This is due to multiple factors, including: a lack of focus from government and funding, discoverability, a lack of community, sheer language complexity, difficulty in reproducing papers and no benchmarks to compare techniques. To begin to address the identified problems, MASAKHANE, an open-source, continent-wide, distributed, online research effort for machine translation for African languages, was founded. In this paper, we discuss our methodology for building the community and spurring research from the African continent, as well as outline the success of the community in terms of addressing the identified problems affecting African NLP.

Nicol Visser, Simon Malan, Danel Slabbert et al.

Pure speech language models aim to learn language directly from raw audio without textual resources. A key challenge is that discrete tokens from self-supervised speech encoders result in excessively long sequences, motivating recent work on syllable-like units. However, methods like Sylber and SyllableLM rely on intricate multi-stage training pipelines. We propose ZeroSyl, a simple training-free method to extract syllable boundaries and embeddings directly from a frozen WavLM model. Using L2 norms of features in WavLM's intermediate layers, ZeroSyl achieves competitive syllable segmentation performance. The resulting segments are mean-pooled, discretized using K-means, and used to train a language model. ZeroSyl outperforms prior syllabic tokenizers across lexical, syntactic, and narrative benchmarks. Scaling experiments show that while finer-grained units are beneficial for lexical tasks, our discovered syllabic units exhibit better scaling behavior for syntactic modeling.

Christiaan Jacobs, Annelien Smith, Daleen Klop et al.

We develop automatic speech recognition (ASR) systems for stories told by Afrikaans and isiXhosa preschool children. Oral narratives provide a way to assess children's language development before they learn to read. We consider a range of prior child-speech ASR strategies to determine which is best suited to this unique setting. Using Whisper and only 5 minutes of transcribed in-domain child speech, we find that additional in-domain adult data (adult speech matching the story domain) provides the biggest improvement, especially when coupled with voice conversion. Semi-supervised learning also helps for both languages, while parameter-efficient fine-tuning helps on Afrikaans but not on isiXhosa (which is under-represented in the Whisper model). Few child-speech studies look at non-English data, and even fewer at the preschool ages of 4 and 5. Our work therefore represents a unique validation of a wide range of previous child-speech ASR strategies in an under-explored setting.

Matthew Baas, Pieter Scholtz, Arnav Mehta et al.

Codec-based text-to-speech (TTS) models have shown impressive quality with zero-shot voice cloning abilities. However, they often struggle with more expressive references or complex text inputs. We present MARS6, a robust encoder-decoder transformer for rapid, expressive TTS. MARS6 is built on recent improvements in spoken language modelling. Utilizing a hierarchical setup for its decoder, new speech tokens are processed at a rate of only 12 Hz, enabling efficient modelling of long-form text while retaining reconstruction quality. We combine several recent training and inference techniques to reduce repetitive generation and improve output stability and quality. This enables the 70M-parameter MARS6 to achieve similar performance to models many times larger. We show this in objective and subjective evaluations, comparing TTS output quality and reference speaker cloning ability. Project page: https://camb-ai.github.io/mars6-turbo/

Herman Kamper, Benjamin van Niekerk

We revisit a self-supervised method that segments unlabelled speech into word-like segments. We start from the two-stage duration-penalised dynamic programming method that performs zero-resource segmentation without learning an explicit lexicon. In the first acoustic unit discovery stage, we replace contrastive predictive coding features with HuBERT. After word segmentation in the second stage, we get an acoustic word embedding for each segment by averaging HuBERT features. These embeddings are clustered using K-means to get a lexicon. The result is good full-coverage segmentation with a lexicon that achieves state-of-the-art performance on the ZeroSpeech benchmarks.

Retief Louw, Emma Sharratt, Febe de Wet et al.

Developing narrative and comprehension skills in early childhood is critical for later literacy. However, teachers in large preschool classrooms struggle to accurately identify students who require intervention. We present a system for automatically assessing oral narratives of preschool children in Afrikaans and isiXhosa. The system uses automatic speech recognition followed by a machine learning scoring model to predict narrative and comprehension scores. For scoring predicted transcripts, we compare a linear model to a large language model (LLM). The LLM-based system outperforms the linear model in most cases, but the linear system is competitive despite its simplicity. The LLM-based system is comparable to a human expert in flagging children who require intervention. We lay the foundation for automatic oral assessments in classrooms, giving teachers extra capacity to focus on personalised support for children's learning.

Emma Sharratt, Annelien Smith, Retief Louw et al.

Oral narrative skills are strong predictors of later literacy development. This study examines the features of oral narratives from children who were identified by experts as requiring intervention. Using simple machine learning methods, we analyse recorded stories from four- and five-year-old Afrikaans- and isiXhosa-speaking children. Consistent with prior research, we identify lexical diversity (unique words) and length-based features (mean utterance length) as indicators of typical development, but features like articulation rate prove less informative. Despite cross-linguistic variation in part-of-speech patterns, the use of specific verbs and auxiliaries associated with goal-directed storytelling is correlated with a reduced likelihood of requiring intervention. Our analysis of two linguistically distinct languages reveals both language-specific and shared predictors of narrative proficiency, with implications for early assessment in multilingual contexts.

Nicol Visser, Herman Kamper

Spoken language models (SLMs) operate on acoustic units obtained by discretizing self-supervised speech representations. Although the characteristics of these units directly affect performance, the interaction between codebook size and unit coarseness (i.e., duration) remains unexplored. We investigate SLM performance as we vary codebook size and unit coarseness using the simple duration-penalized dynamic programming (DPDP) method. New analyses are performed across different linguistic levels. At the phone and word levels, coarseness provides little benefit, as long as the codebook size is chosen appropriately. However, when producing whole sentences in a resynthesis task, SLMs perform better with coarser units. In lexical and syntactic language modeling tasks, coarser units also give higher accuracies at lower bitrates. We therefore show that coarser units aren't always better, but that DPDP is a simple and efficient way to obtain coarser units for the tasks where they are beneficial.

Leanne Nortje, Dan Oneaţă, Yevgen Matusevych et al.

When children learn new words, they employ constraints such as the mutual exclusivity (ME) bias: a novel word is mapped to a novel object rather than a familiar one. This bias has been studied computationally, but only in models that use discrete word representations as input, ignoring the high variability of spoken words. We investigate the ME bias in the context of visually grounded speech models that learn from natural images and continuous speech audio. Concretely, we train a model on familiar words and test its ME bias by asking it to select between a novel and a familiar object when queried with a novel word. To simulate prior acoustic and visual knowledge, we experiment with several initialisation strategies using pretrained speech and vision networks. Our findings reveal the ME bias across the different initialisation approaches, with a stronger bias in models with more prior (in particular, visual) knowledge. Additional tests confirm the robustness of our results, even when different loss functions are considered.

Danel Adendorff, Simon Malan, Herman Kamper

Zero-resource word segmentation and clustering systems aim to tokenise speech into word-like units without access to text labels. Despite progress, the induced lexicons are still far from perfect. In an idealised setting with gold word boundaries, we ask whether performance is limited by the representation of word segments, or by the clustering methods that group them into word-like types. We combine a range of self-supervised speech features (continuous/discrete, frame/word-level) with different clustering methods (K-means, hierarchical, graph-based) on English and Mandarin data. The best system uses graph clustering with dynamic time warping on continuous features. Faster alternatives use graph clustering with cosine distance on averaged continuous features or edit distance on discrete unit sequences. Through controlled experiments that isolate either the representations or the clustering method, we demonstrate that representation variability across segments of the same word type -- rather than clustering -- is the primary factor limiting performance.

Reuben Smit, Retief Louw, Herman Kamper

We explore an ASR-free method for isolated word reading assessment in low-resource settings. Our few-shot approach compares input child speech to a small set of adult-provided reference templates. Inputs and templates are encoded using intermediate layers from large self-supervised learned (SSL) models. Using an Afrikaans child speech benchmark, we investigate design options such as discretising SSL features and barycentre averaging of the templates. Idealised experiments show reasonable performance for adults, but a substantial drop for child speech input, even with child templates. Despite the success of employing SSL representations in low-resource speech tasks, our work highlights the limitations of SSL representations for processing child data when used in a few-shot classification system.

Herman Kamper, Benjamin van Niekerk, Julian Zaïdi et al.

We introduce LinearVC, a simple voice conversion method that sheds light on the structure of self-supervised representations. First, we show that simple linear transformations of self-supervised features effectively convert voices. Next, we probe the geometry of the feature space by constraining the set of allowed transformations. We find that just rotating the features is sufficient for high-quality voice conversion. This suggests that content information is embedded in a low-dimensional subspace which can be linearly transformed to produce a target voice. To validate this hypothesis, we finally propose a method that explicitly factorizes content and speaker information using singular value decomposition; the resulting linear projection with a rank of just 100 gives competitive conversion results. Our work has implications for both practical voice conversion and a broader understanding of self-supervised speech representations. Samples and code: https://www.kamperh.com/linearvc/.

Dan Oneata, Herman Kamper

Visually grounded speech models link speech to images. We extend this connection by linking images to text via an existing image captioning system, and as a result gain the ability to map speech audio directly to text. This approach can be used for speech translation with just images by having the audio in a different language from the generated captions. We investigate such a system on a real low-resource language, Yorùbá, and propose a Yorùbá-to-English speech translation model that leverages pretrained components in order to be able to learn in the low-resource regime. To limit overfitting, we find that it is essential to use a decoding scheme that produces diverse image captions for training. Results show that the predicted translations capture the main semantics of the spoken audio, albeit in a simpler and shorter form.

Matthew Baas, Benjamin van Niekerk, Herman Kamper

Any-to-any voice conversion aims to transform source speech into a target voice with just a few examples of the target speaker as a reference. Recent methods produce convincing conversions, but at the cost of increased complexity -- making results difficult to reproduce and build on. Instead, we keep it simple. We propose k-nearest neighbors voice conversion (kNN-VC): a straightforward yet effective method for any-to-any conversion. First, we extract self-supervised representations of the source and reference speech. To convert to the target speaker, we replace each frame of the source representation with its nearest neighbor in the reference. Finally, a pretrained vocoder synthesizes audio from the converted representation. Objective and subjective evaluations show that kNN-VC improves speaker similarity with similar intelligibility scores to existing methods. Code, samples, trained models: https://bshall.github.io/knn-vc

Leanne Nortje, Benjamin van Niekerk, Herman Kamper

We propose a visually grounded speech model that acquires new words and their visual depictions from just a few word-image example pairs. Given a set of test images and a spoken query, we ask the model which image depicts the query word. Previous work has simplified this problem by either using an artificial setting with digit word-image pairs or by using a large number of examples per class. We propose an approach that can work on natural word-image pairs but with less examples, i.e. fewer shots. Our approach involves using the given word-image example pairs to mine new unsupervised word-image training pairs from large collections of unlabelled speech and images. Additionally, we use a word-to-image attention mechanism to determine word-image similarity. With this new model, we achieve better performance with fewer shots than any existing approach.

Ruan van der Merwe, Herman Kamper

We consider the problem of few-shot spoken word classification in a setting where a model is incrementally introduced to new word classes. This would occur in a user-defined keyword system where new words can be added as the system is used. In such a continual learning scenario, a model might start to misclassify earlier words as newer classes are added, i.e. catastrophic forgetting. To address this, we propose an extension to model-agnostic meta-learning (MAML): each inner learning loop, where a model "learns how to learn'' new classes, ends with a single gradient update using stored templates from all the classes that the model has already seen (one template per class). We compare this method to OML (another extension of MAML) in few-shot isolated-word classification experiments on Google Commands and FACC. Our method consistently outperforms OML in experiments where the number of shots and the final number of classes are varied.

Herman Kamper

Recent work on unsupervised speech segmentation has used self-supervised models with phone and word segmentation modules that are trained jointly. This paper instead revisits an older approach to word segmentation: bottom-up phone-like unit discovery is performed first, and symbolic word segmentation is then performed on top of the discovered units (without influencing the lower level). To do this, I propose a new unit discovery model, a new symbolic word segmentation model, and then chain the two models to segment speech. Both models use dynamic programming to minimize segment costs from a self-supervised network with an additional duration penalty that encourages longer units. Concretely, for acoustic unit discovery, duration-penalized dynamic programming (DPDP) is used with a contrastive predictive coding model as the scoring network. For word segmentation, DPDP is applied with an autoencoding recurrent neural as the scoring network. The two models are chained in order to segment speech. This approach gives comparable word segmentation results to state-of-the-art joint self-supervised segmentation models on an English benchmark. On French, Mandarin, German and Wolof data, it outperforms previous systems on the ZeroSpeech benchmarks. Analysis shows that the chained DPDP system segments shorter filler words well, but longer words might require some external top-down signal.

Kayode Olaleye, Dan Oneata, Herman Kamper

Keyword localisation is the task of finding where in a speech utterance a given query keyword occurs. We investigate to what extent keyword localisation is possible using a visually grounded speech (VGS) model. VGS models are trained on unlabelled images paired with spoken captions. These models are therefore self-supervised -- trained without any explicit textual label or location information. To obtain training targets, we first tag training images with soft text labels using a pretrained visual classifier with a fixed vocabulary. This enables a VGS model to predict the presence of a written keyword in an utterance, but not its location. We consider four ways to equip VGS models with localisations capabilities. Two of these -- a saliency approach and input masking -- can be applied to an arbitrary prediction model after training, while the other two -- attention and a score aggregation approach -- are incorporated directly into the structure of the model. Masked-based localisation gives some of the best reported localisation scores from a VGS model, with an accuracy of 57% when the system knows that a keyword occurs in an utterance and need to predict its location. In a setting where localisation is performed after detection, an $F_1$ of 25% is achieved, and in a setting where a keyword spotting ranking pass is first performed, we get a localisation P@10 of 32%. While these scores are modest compared to the idealised setting with unordered bag-of-word-supervision (from transcriptions), these models do not receive any textual or location supervision. Further analyses show that these models are limited by the first detection or ranking pass. Moreover, individual keyword localisation performance is correlated with the tagging performance from the visual classifier. We also show qualitatively how and where semantic mistakes occur, e.g. that the model locates surfer when queried with ocean.

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.

Matthew Baas, Herman Kamper

Voice conversion (VC) could be used to improve speech recognition systems in low-resource languages by using it to augment limited training data. However, VC has not been widely used for this purpose because of practical issues such as compute speed and limitations when converting to and from unseen speakers. Moreover, it is still unclear whether a VC model trained on one well-resourced language can be applied to speech from another low-resource language for the aim of data augmentation. In this work we assess whether a VC system can be used cross-lingually to improve low-resource speech recognition. We combine several recent techniques to design and train a practical VC system in English, and then use this system to augment data for training speech recognition models in several low-resource languages. When using a sensible amount of VC augmented data, speech recognition performance is improved in all four low-resource languages considered. We also show that VC-based augmentation is superior to SpecAugment (a widely used signal processing augmentation method) in the low-resource languages considered.

Benjamin van Niekerk, Leanne Nortje, Matthew Baas et al.

Contrastive predictive coding (CPC) aims to learn representations of speech by distinguishing future observations from a set of negative examples. Previous work has shown that linear classifiers trained on CPC features can accurately predict speaker and phone labels. However, it is unclear how the features actually capture speaker and phonetic information, and whether it is possible to normalize out the irrelevant details (depending on the downstream task). In this paper, we first show that the per-utterance mean of CPC features captures speaker information to a large extent. Concretely, we find that comparing means performs well on a speaker verification task. Next, probing experiments show that standardizing the features effectively removes speaker information. Based on this observation, we propose a speaker normalization step to improve acoustic unit discovery using K-means clustering of CPC features. Finally, we show that a language model trained on the resulting units achieves some of the best results in the ZeroSpeech2021~Challenge.

Christiaan Jacobs, Herman Kamper

Acoustic word embedding models map variable duration speech segments to fixed dimensional vectors, enabling efficient speech search and discovery. Previous work explored how embeddings can be obtained in zero-resource settings where no labelled data is available in the target language. The current best approach uses transfer learning: a single supervised multilingual model is trained using labelled data from multiple well-resourced languages and then applied to a target zero-resource language (without fine-tuning). However, it is still unclear how the specific choice of training languages affect downstream performance. Concretely, here we ask whether it is beneficial to use training languages related to the target. Using data from eleven languages spoken in Southern Africa, we experiment with adding data from different language families while controlling for the amount of data per language. In word discrimination and query-by-example search evaluations, we show that training on languages from the same family gives large improvements. Through finer-grained analysis, we show that training on even just a single related language gives the largest gain. We also find that adding data from unrelated languages generally doesn't hurt performance.

Kayode Olaleye, Herman Kamper

Visually grounded speech models learn from images paired with spoken captions. By tagging images with soft text labels using a trained visual classifier with a fixed vocabulary, previous work has shown that it is possible to train a model that can detect whether a particular text keyword occurs in speech utterances or not. Here we investigate whether visually grounded speech models can also do keyword localisation: predicting where, within an utterance, a given textual keyword occurs without any explicit text-based or alignment supervision. We specifically consider whether incorporating attention into a convolutional model is beneficial for localisation. Although absolute localisation performance with visually supervised models is still modest (compared to using unordered bag-of-word text labels for supervision), we show that attention provides a large gain in performance over previous visually grounded models. As in many other speech-image studies, we find that many of the incorrect localisations are due to semantic confusions, e.g. locating the word 'backstroke' for the query keyword 'swimming'.

Matthew Baas, Herman Kamper

Voice conversion is the task of converting a spoken utterance from a source speaker so that it appears to be said by a different target speaker while retaining the linguistic content of the utterance. Recent advances have led to major improvements in the quality of voice conversion systems. However, to be useful in a wider range of contexts, voice conversion systems would need to be (i) trainable without access to parallel data, (ii) work in a zero-shot setting where both the source and target speakers are unseen during training, and (iii) run in real time or faster. Recent techniques fulfil one or two of these requirements, but not all three. This paper extends recent voice conversion models based on generative adversarial networks (GANs), to satisfy all three of these conditions. We specifically extend the recent StarGAN-VC model by conditioning it on a speaker embedding (from a potentially unseen speaker). This allows the model to be used in a zero-shot setting, and we therefore call it StarGAN-ZSVC. We compare StarGAN-ZSVC against other voice conversion techniques in a low-resource setting using a small 9-minute training set. Compared to AutoVC -- another recent neural zero-shot approach -- we observe that StarGAN-ZSVC gives small improvements in the zero-shot setting, showing that real-time zero-shot voice conversion is possible even for a model trained on very little data. Further work is required to see whether scaling up StarGAN-ZSVC will also improve zero-shot voice conversion quality in high-resource contexts.

Christiaan Jacobs, Yevgen Matusevych, Herman Kamper

Acoustic word embeddings (AWEs) are fixed-dimensional representations of variable-length speech segments. For zero-resource languages where labelled data is not available, one AWE approach is to use unsupervised autoencoder-based recurrent models. Another recent approach is to use multilingual transfer: a supervised AWE model is trained on several well-resourced languages and then applied to an unseen zero-resource language. We consider how a recent contrastive learning loss can be used in both the purely unsupervised and multilingual transfer settings. Firstly, we show that terms from an unsupervised term discovery system can be used for contrastive self-supervision, resulting in improvements over previous unsupervised monolingual AWE models. Secondly, we consider how multilingual AWE models can be adapted to a specific zero-resource language using discovered terms. We find that self-supervised contrastive adaptation outperforms adapted multilingual correspondence autoencoder and Siamese AWE models, giving the best overall results in a word discrimination task on six zero-resource languages.

Yevgen Matusevych, Herman Kamper, Thomas Schatz et al.

Non-native speakers show difficulties with spoken word processing. Many studies attribute these difficulties to imprecise phonological encoding of words in the lexical memory. We test an alternative hypothesis: that some of these difficulties can arise from the non-native speakers' phonetic perception. We train a computational model of phonetic learning, which has no access to phonology, on either one or two languages. We first show that the model exhibits predictable behaviors on phone-level and word-level discrimination tasks. We then test the model on a spoken word processing task, showing that phonology may not be necessary to explain some of the word processing effects observed in non-native speakers. We run an additional analysis of the model's lexical representation space, showing that the two training languages are not fully separated in that space, similarly to the languages of a bilingual human speaker.

Herman Kamper, Benjamin van Niekerk

We investigate segmenting and clustering speech into low-bitrate phone-like sequences without supervision. We specifically constrain pretrained self-supervised vector-quantized (VQ) neural networks so that blocks of contiguous feature vectors are assigned to the same code, thereby giving a variable-rate segmentation of the speech into discrete units. Two segmentation methods are considered. In the first, features are greedily merged until a prespecified number of segments are reached. The second uses dynamic programming to optimize a squared error with a penalty term to encourage fewer but longer segments. We show that these VQ segmentation methods can be used without alteration across a wide range of tasks: unsupervised phone segmentation, ABX phone discrimination, same-different word discrimination, and as inputs to a symbolic word segmentation algorithm. The penalized dynamic programming method generally performs best. While performance on individual tasks is only comparable to the state-of-the-art in some cases, in all tasks a reasonable competing approach is outperformed at a substantially lower bitrate.

Kayode Olaleye, Benjamin van Niekerk, Herman Kamper

Developments in weakly supervised and self-supervised models could enable speech technology in low-resource settings where full transcriptions are not available. We consider whether keyword localisation is possible using two forms of weak supervision where location information is not provided explicitly. In the first, only the presence or absence of a word is indicated, i.e. a bag-of-words (BoW) labelling. In the second, visual context is provided in the form of an image paired with an unlabelled utterance; a model then needs to be trained in a self-supervised fashion using the paired data. For keyword localisation, we adapt a saliency-based method typically used in the vision domain. We compare this to an existing technique that performs localisation as a part of the network architecture. While the saliency-based method is more flexible (it can be applied without architectural restrictions), we identify a critical limitation when using it for keyword localisation. Of the two forms of supervision, the visually trained model performs worse than the BoW-trained model. We show qualitatively that the visually trained model sometimes locate semantically related words, but this is not consistent. While our results show that there is some signal allowing for localisation, it also calls for other localisation methods better matched to these forms of weak supervision.

Lisa van Staden, Herman Kamper

Many speech processing tasks involve measuring the acoustic similarity between speech segments. Acoustic word embeddings (AWE) allow for efficient comparisons by mapping speech segments of arbitrary duration to fixed-dimensional vectors. For zero-resource speech processing, where unlabelled speech is the only available resource, some of the best AWE approaches rely on weak top-down constraints in the form of automatically discovered word-like segments. Rather than learning embeddings at the segment level, another line of zero-resource research has looked at representation learning at the short-time frame level. Recent approaches include self-supervised predictive coding and correspondence autoencoder (CAE) models. In this paper we consider whether these frame-level features are beneficial when used as inputs for training to an unsupervised AWE model. We compare frame-level features from contrastive predictive coding (CPC), autoregressive predictive coding and a CAE to conventional MFCCs. These are used as inputs to a recurrent CAE-based AWE model. In a word discrimination task on English and Xitsonga data, all three representation learning approaches outperform MFCCs, with CPC consistently showing the biggest improvement. In cross-lingual experiments we find that CPC features trained on English can also be transferred to Xitsonga.

Leanne Nortje, Herman Kamper

We propose direct multimodal few-shot models that learn a shared embedding space of spoken words and images from only a few paired examples. Imagine an agent is shown an image along with a spoken word describing the object in the picture, e.g. pen, book and eraser. After observing a few paired examples of each class, the model is asked to identify the "book" in a set of unseen pictures. Previous work used a two-step indirect approach relying on learned unimodal representations: speech-speech and image-image comparisons are performed across the support set of given speech-image pairs. We propose two direct models which instead learn a single multimodal space where inputs from different modalities are directly comparable: a multimodal triplet network (MTriplet) and a multimodal correspondence autoencoder (MCAE). To train these direct models, we mine speech-image pairs: the support set is used to pair up unlabelled in-domain speech and images. In a speech-to-image digit matching task, direct models outperform indirect models, with the MTriplet achieving the best multimodal five-shot accuracy. We show that the improvements are due to the combination of unsupervised and transfer learning in the direct models, and the absence of two-step compounding errors.