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/

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.

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.

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.

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 letter 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.

Abstract 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 initialization strategies using pretrained speech and vision networks. Our findings reveal the ME bias across the different initialization 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. Based on detailed analyses to piece out the model’s representation space, we attribute the ME bias to how familiar and novel classes are distinctly separated in the resulting space.

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/.

We propose a visually grounded speech model that learns new words and their visual depictions from just a few wordimage 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, Yorùá. 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 Yorùá show the benefit of transferring knowledge from a multimodal model trained on a larger set of English speech-image data.1

We propose a visually grounded speech model that acquires new words and their visual depictions from just a few wordimage 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.

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.

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.

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 learned by predicting a distribution over the discrete units. By modeling uncertainty, soft units capture more content information, improving the intelligibility and naturalness of converted speech.12

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.

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.

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/.

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.

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 Yoruba=-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 Yoruba utterances. This enables cross-lingual keyword localisation: a written English query is detected and located in Yoruba 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.

Yevgen Matusevych, Herman Kamper, Thomas Schatz, Naomi Feldman, Sharon Goldwater. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

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.

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.

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.

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.

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'.

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 re-current 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.

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.

Acoustic word embeddings are fixed-dimensional representations of variable-length speech segments. Such embeddings can form the basis for speech search, indexing and discovery systems when conventional speech recognition is not possible. In zero-resource settings where unlabelled speech is the only available resource, we need a method that gives robust embeddings on an arbitrary language. Here we explore multilingual transfer: we train a single supervised embedding model on labelled data from multiple well-resourced languages and then apply it to unseen zero-resource languages. We consider three multilingual recurrent neural network (RNN) models: a classifier trained on the joint vocabularies of all training languages; a Siamese RNN trained to discriminate between same and different words from multiple languages; and a correspondence autoencoder (CAE) RNN trained to reconstruct word pairs. In a word discrimination task on six target languages, all of these models outperform state-of-the-art unsupervised models trained on the zero-resource languages themselves, giving relative improvements of more than 30% in average precision. When using only a few training languages, the multilingual CAE-RNN performs better, but with more training languages the other multilingual models perform similarly. Using more training languages is generally beneficial, but improvements are marginal on some languages. We present probing experiments which show that the CAE-RNN encodes more phonetic, word duration, language identity and speaker information than the other multilingual models.

Given a large amount of unannotated speech in a low-resource language, can we classify the speech utterances by topicƒ We consider this question in the setting where a small amount of speech in the low-resource language is paired with text translations in a high-resource language. We develop an effective cross-lingual topic classifier by training on just 20 hours of translated speech, using a recent model for direct speech-to-text translation. While the translations are poor, they are still good enough to correctly classify the topic of 1-minute speech segments over 70% of the time—a 20% improvement over a majority-class baseline. Such a system could be useful for humanitarian applications like crisis response, where incoming speech in a foreign low-resource language must be quickly assessed for further action.

Acoustic word embeddings are fixed-dimensional representations of variable-length speech segments. In settings where unlabelled speech is the only available resource, such embeddings can be used in "zero-resource" speech search, indexing and discovery systems. Here we propose to train a single supervised embedding model on labelled data from multiple well-resourced languages and then apply it to unseen zero-resource languages. For this transfer learning approach, we consider two multilingual recurrent neural network models: a discriminative classifier trained on the joint vocabularies of all training languages, and a correspondence autoencoder trained to reconstruct word pairs. We test these using a word discrimination task on six target zero-resource languages. When trained on seven well-resourced languages, both models perform similarly and outperform unsupervised models trained on the zero-resource languages. With just a single training language, the second model works better, but performance depends more on the particular training-testing language pair.

In this paper, we explore vector quantization for acoustic unit discovery. Leveraging unlabelled data, we aim to learn discrete representations of speech that separate phonetic content from speaker-specific details. We propose two neural models to tackle this challenge - both use vector quantization to map continuous features to a finite set of codes. The first model is a type of vector-quantized variational autoencoder (VQ-VAE). The VQ-VAE encodes speech into a sequence of discrete units before reconstructing the audio waveform. Our second model combines vector quantization with contrastive predictive coding (VQ-CPC). The idea is to learn a representation of speech by predicting future acoustic units. We evaluate the models on English and Indonesian data for the ZeroSpeech 2020 challenge. In ABX phone discrimination tests, both models outperform all submissions to the 2019 and 2020 challenges, with a relative improvement of more than 30%. The models also perform competitively on a downstream voice conversion task. Of the two, VQ-CPC performs slightly better in general and is simpler and faster to train. Finally, probing experiments show that vector quantization is an effective bottleneck, forcing the models to discard speaker information.

We consider the task of multimodal one-shot speech-image matching. An agent is shown a picture along with a spoken word describing the object in the picture, e.g. cookie, broccoli and After observing one paired speech-image example per class, it is shown a new set of unseen pictures, and asked to pick the ice-cream. Previous work attempted to tackle this problem using transfer learning: supervised models are trained on labelled background data not containing any of the one-shot classes. Here we compare transfer learning to unsupervised models trained on unlabelled in-domain data. On a dataset of paired isolated spoken and visual digits, we specifically compare unsupervised autoencoder-like models to supervised classifier and Siamese neural networks. In both unimodal and multimodal few-shot matching experiments, we find that transfer learning outperforms unsupervised training. We also present experiments towards combining the two methodologies, but find that transfer learning still performs best (despite idealised experiments showing the benefits of unsupervised learning).

In zero-resource settings where transcribed speech audio is unavailable, unsupervised feature learning is essential for downstream speech processing tasks. Here we compare two recent methods for frame-level acoustic feature learning. For both methods, unsupervised term discovery is used to find pairs of word examples of the same unknown type. Dynamic programming is then used to align the feature frames between each word pair, serving as weak top-down supervision for the two models. For the correspondence autoencoder (CAE), matching frames are presented as input-output pairs. The Triamese network uses a contrastive loss to reduce the distance between frames of the same predicted word type while increasing the distance between negative examples. For the first time, these feature extractors are compared on the same discrimination tasks using the same weak supervision pairs. We find that, on the two datasets considered here, the CAE outperforms the Triamese network. However, we show that a new hybrid correspondence-Triamese approach (CTriamese), consistently outperforms both the CAE and Triamese models in terms of average precision and ABX error rates on both English and Xitsonga evaluation data.

Imagine a robot is shown new concepts visually together with spoken tags, e.g. "milk", "eggs", "butter". After seeing one paired audiovisual example per class, it is shown a new set of unseen instances of these objects, and asked to pick the "milk". Without receiving any hard labels, could it learn to match the new continuous speech input to the correct visual instance? Although unimodal one-shot learning has been studied, where one labelled example in a single modality is given per class, this example motivates multimodal one-shot learning. Our main contribution is to formally define this task, and to propose several baseline and advanced models. We use a dataset of paired spoken and visual digits to specifically investigate recent advances in Siamese convolutional neural networks. Our best Siamese model achieves twice the accuracy of a nearest neighbour model using pixel-distance over images and dynamic time warping over speech in 11-way cross-modal matching.

We investigate unsupervised models that can map a variable-duration speech segment to a fixed-dimensional representation. In settings where unlabelled speech is the only available resource, such acoustic word embeddings can form the basis for "zero-resource" speech search, discovery and indexing systems. Most existing unsupervised embedding methods still use some supervision, such as word or phoneme boundaries. Here we propose the encoder-decoder correspondence autoencoder (EncDec-CAE), which, instead of true word segments, uses automatically discovered segments: an unsupervised term discovery system finds pairs of words of the same unknown type, and the EncDec-CAE is trained to reconstruct one word given the other as input. We compare it to a standard encoder-decoder autoencoder (AE), a variational AE with a prior over its latent embedding, and downsampling. EncDec-CAE outperforms its closest competitor by 29% relative in average precision on two languages in a word discrimination task.

A number of recent studies have started to investigate how speech systems can be trained on untranscribed speech by leveraging accompanying images at training time. Examples of tasks include keyword prediction and within- and across-mode retrieval. Here we consider how such models can be used for query-by-example (QbE) search, the task of retrieving utterances relevant to a given spoken query. We are particularly interested in semantic QbE, where the task is not only to retrieve utterances containing exact instances of the query, but also utterances whose meaning is relevant to the query. We follow a segmental QbE approach where variable-duration speech segments (queries, search utterances) are mapped to fixed-dimensional embedding vectors. We show that a QbE system using an embedding function trained on visually grounded speech data outperforms a purely acoustic QbE system in terms of both exact and semantic retrieval performance.

For our submission to the ZeroSpeech 2019 challenge, we apply discrete latent-variable neural networks to unlabelled speech and use the discovered units for speech synthesis. Unsupervised discrete subword modelling could be useful for studies of phonetic category learning in infants or in low-resource speech technology requiring symbolic input. We use an autoencoder (AE) architecture with intermediate discretisation. We decouple acoustic unit discovery from speaker modelling by conditioning the AE's decoder on the training speaker identity. At test time, unit discovery is performed on speech from an unseen speaker, followed by unit decoding conditioned on a known target speaker to obtain reconstructed filterbanks. This output is fed to a neural vocoder to synthesise speech in the target speaker's voice. For discretisation, categorical variational autoencoders (CatVAEs), vector-quantised VAEs (VQ-VAEs) and straight-through estimation are compared at different compression levels on two languages. Our final model uses convolutional encoding, VQ-VAE discretisation, deconvolutional decoding and an FFTNet vocoder. We show that decoupled speaker conditioning intrinsically improves discrete acoustic representations, yielding competitive synthesis quality compared to the challenge baseline.

We compare features for dynamic time warping (DTW) when used to bootstrap keyword spotting (KWS) in an almost zero-resource setting. Such quickly-deployable systems aim to support United Nations (UN) humanitarian relief efforts in parts of Africa with severely under-resourced languages. Our objective is to identify acoustic features that provide acceptable KWS performance in such environments. As supervised resource, we restrict ourselves to a small, easily acquired and independently compiled set of isolated keywords. For feature extraction, a multilingual bottleneck feature (BNF) extractor, trained on well-resourced out-of-domain languages, is integrated with a correspondence autoencoder (CAE) trained on extremely sparse in-domain data. On their own, BNFs and CAE features are shown to achieve a more than 2% absolute performance improvement over baseline MFCCs. However, by using BNFs as input to the CAE, even better performance is achieved, with a more than 11% absolute improvement in ROC AUC over MFCCs and more than twice as many top-10 retrievals for two evaluated languages, English and Luganda. We conclude that integrating BNFs with the CAE allows both large out-of-domain and sparse in-domain resources to be exploited for improved ASR-free keyword spotting.

Recent work has shown that speech paired with images can be used to learn semantically meaningful speech representations even without any textual supervision. In real-world low-resource settings, however, we often have access to some transcribed speech. We study whether and how visual grounding is useful in the presence of varying amounts of textual supervision. In particular, we consider the task of semantic speech retrieval in a low-resource setting. We use a previously studied data set and task, where models are trained on images with spoken captions and evaluated on human judgments of semantic relevance. We propose a multitask learning approach to leverage both visual and textual modalities, with visual supervision in the form of keyword probabilities from an external tagger. We find that visual grounding is helpful even in the presence of textual supervision, and we analyze this effect over a range of sizes of transcribed data sets. With ~5 hours of transcribed speech, we obtain 23% higher average precision when also using visual supervision.

There is growing interest in models that can learn from unlabelled speech paired with visual context. This setting is relevant for low-resource speech processing, robotics, and human language acquisition research. Here we study how a visually grounded speech model, trained on images of scenes paired with spoken captions, captures aspects of semantics. We use an external image tagger to generate soft text labels from images, which serve as targets for a neural model that maps untranscribed speech to (semantic) keyword labels. We introduce a newly collected data set of human semantic relevance judgements and an associated task, semantic speech retrieval, where the goal is to search for spoken utterances that are semantically relevant to a given text query. Without seeing any text, the model trained on parallel speech and images achieves a precision of almost 60% on its top ten semantic retrievals. Compared to a supervised model trained on transcriptions, our model matches human judgements better by some measures, especially in retrieving non-verbatim semantic matches. We perform an extensive analysis of the model and its resulting representations.

Identifying and grouping the frequently occurring word-like patterns from raw acoustic waveforms is an important task in the zero resource speech processing. Embedded segmental K-means (ES-KMeans) discovers both the word boundaries and the word types from raw data. Starting from an initial set of subword boundaries, the ES-Kmeans iteratively eliminates some of the boundaries to arrive at frequently occurring longer word patterns. Notice that the initial word boundaries will not be adjusted during the process. As a result, the performance of the ES-Kmeans critically depends on the initial subword boundaries. Originally, syllable boundaries were used to initialize ES-Kmeans. In this paper, we propose to use a phoneme segmentation method that produces boundaries closer to true boundaries for ES-KMeans initialization. The use of shorter units increases the number of initial boundaries which leads to a significant increment in the computational complexity. To reduce the computational cost, we extract compact lower dimensional embeddings from an auto-encoder. The proposed algorithm is benchmarked on Zero Resource 2017 challenge, which consists of 70 hours of unlabeled data across three languages, viz. English, French, and Mandarin. The proposed algorithm outperforms the baseline system without any language-specific parameter tuning.

Denoising autoencoders (DAEs) have proven useful for unsupervised representation learning, but a thorough theoretical understanding is still lacking of how the input noise influences learning. Here we develop theory for how noise influences learning in DAEs. By focusing on linear DAEs, we are able to derive analytic expressions that exactly describe their learning dynamics. We verify our theoretical predictions with simulations as well as experiments on MNIST and CIFAR-10. The theory illustrates how, when tuned correctly, noise allows DAEs to ignore low variance directions in the inputs while learning to reconstruct them. Furthermore, in a comparison of the learning dynamics of DAEs to standard regularised autoencoders, we show that noise has a similar regularisation effect to weight decay, but with faster training dynamics. We also show that our theoretical predictions approximate learning dynamics on real-world data and qualitatively match observed dynamics in nonlinear DAEs.

Speech-to-text translation has many potential applications for low-resource languages, but the typical approach of cascading speech recognition with machine translation is often impossible, since the transcripts needed to train a speech recognizer are usually not available for low-resource languages. Recent work has found that neural encoder-decoder models can learn to directly translate foreign speech in high-resource scenarios, without the need for intermediate transcription. We investigate whether this approach also works in settings where both data and computation are limited. To make the approach efficient, we make several architectural changes, including a change from character-level to word-level decoding. We find that this choice yields crucial speed improvements that allow us to train with fewer computational resources, yet still performs well on frequent words. We explore models trained on between 20 and 160 hours of data, and find that although models trained on less data have considerably lower BLEU scores, they can still predict words with relatively high precision and recall---around 50% for a model trained on 50 hours of data, versus around 60% for the full 160 hour model. Thus, they may still be useful for some low-resource scenarios.

We use dynamic time warping (DTW) as supervision for training a convolutional neural network (CNN) based keyword spotting system using a small set of spoken isolated keywords. The aim is to allow rapid deployment of a keyword spotting system in a new language to support urgent United Nations (UN) relief programmes in parts of Africa where languages are extremely under-resourced and the development of annotated speech resources is infeasible. First, we use 1920 recorded keywords (40 keyword types, 34 minutes of speech) as exemplars in a DTW-based template matching system and apply it to untranscribed broadcast speech. Then, we use the resulting DTW scores as targets to train a CNN on the same unlabelled speech. In this way we use just 34 minutes of labelled speech, but leverage a large amount of unlabelled data for training. While the resulting CNN keyword spotter cannot match the performance of the DTW-based system, it substantially outperforms a CNN classifier trained only on the keywords, improving the area under the ROC curve from 0.54 to 0.64. Because our CNN system is several orders of magnitude faster at runtime than the DTW system, it represents the most viable keyword spotter on this extremely limited dataset.

Stochastic regularisation is an important weapon in the arsenal of a deep learning practitioner. However, despite recent theoretical advances, our understanding of how noise influences signal propagation in deep neural networks remains limited. By extending recent work based on mean field theory, we develop a new framework for signal propagation in stochastic regularised neural networks. Our \textit{noisy signal propagation} theory can incorporate several common noise distributions, including additive and multiplicative Gaussian noise as well as dropout. We use this framework to investigate initialisation strategies for noisy ReLU networks. We show that no critical initialisation strategy exists using additive noise, with signal propagation exploding regardless of the selected noise distribution. For multiplicative noise (e.g.\ dropout), we identify alternative critical initialisation strategies that depend on the second moment of the noise distribution. Simulations and experiments on real-world data confirm that our proposed initialisation is able to stably propagate signals in deep networks, while using an initialisation disregarding noise fails to do so. Furthermore, we analyse correlation dynamics between inputs. Stronger noise regularisation is shown to reduce the depth to which discriminatory information about the inputs to a noisy ReLU network is able to propagate, even when initialised at criticality. We support our theoretical predictions for these trainable depths with simulations, as well as with experiments on MNIST and CIFAR-10.

Unsupervised segmentation and clustering of unlabelled speech are core problems in zero-resource speech processing. Most approaches lie at methodological extremes: some use probabilistic Bayesian models with convergence guarantees, while others opt for more efficient heuristic techniques. Despite competitive performance in previous work, the full Bayesian approach is difficult to scale to large speech corpora. We introduce an approximation to a recent Bayesian model that still has a clear objective function but improves efficiency by using hard clustering and segmentation rather than full Bayesian inference. Like its Bayesian counterpart, this embedded segmental K-means model (ES-KMeans) represents arbitrary-length word segments as fixed-dimensional acoustic word embeddings. We first compare ES-KMeans to previous approaches on common English and Xitsonga data sets (5 and 2.5 hours of speech): ES-KMeans outperforms a leading heuristic method in word segmentation, giving similar scores to the Bayesian model while being 5 times faster with fewer hyperparameters. However, its clusters are less pure than those of the other models. We then show that ES-KMeans scales to larger corpora by applying it to the 5 languages of the Zero Resource Speech Challenge 2017 (up to 45 hours), where it performs competitively compared to the challenge baseline.

Recent work on unsupervised term discovery (UTD) aims to identify and cluster repeated word-like units from audio alone. These systems are promising for some very low-resource languages where transcribed audio is unavailable, or where no written form of the language exists. However, in some cases it may still be feasible (e.g., through crowdsourcing) to obtain (possibly noisy) text translations of the audio. If so, this information could be used as a source of side information to improve UTD. Here, we present a simple method for rescoring the output of a UTD system using text translations, and test it on a corpus of Spanish audio with English translations. We show that it greatly improves the average precision of the results over a wide range of system configurations and data preprocessing methods.

During language acquisition, infants have the benefit of visual cues to ground spoken language. Robots similarly have access to audio and visual sensors. Recent work has shown that images and spoken captions can be mapped into a meaningful common space, allowing images to be retrieved using speech and vice versa. In this setting of images paired with untranscribed spoken captions, we consider whether computer vision systems can be used to obtain textual labels for the speech. Concretely, we use an image-to-words multi-label visual classifier to tag images with soft textual labels, and then train a neural network to map from the speech to these soft targets. We show that the resulting speech system is able to predict which words occur in an utterance---acting as a spoken bag-of-words classifier---without seeing any parallel speech and text. We find that the model often confuses semantically related words, e.g. "man" and "person", making it even more effective as a semantic keyword spotter.

Query-by-example search often uses dynamic time warping (DTW) for comparing queries and proposed matching segments. Recent work has shown that comparing speech segments by representing them as fixed-dimensional vectors --- acoustic word embeddings --- and measuring their vector distance (e.g., cosine distance) can discriminate between words more accurately than DTW-based approaches. We consider an approach to query-by-example search that embeds both the query and database segments according to a neural model, followed by nearest-neighbor search to find the matching segments. Earlier work on embedding-based query-by-example, using template-based acoustic word embeddings, achieved competitive performance. We find that our embeddings, based on recurrent neural networks trained to optimize word discrimination, achieve substantial improvements in performance and run-time efficiency over the previous approaches.

Recent studies have been revisiting whole words as the basic modelling unit in speech recognition and query applications, instead of phonetic units. Such whole-word segmental systems rely on a function that maps a variable-length speech segment to a vector in a fixed-dimensional space; the resulting acoustic word embeddings need to allow for accurate discrimination between different word types, directly in the embedding space. We compare several old and new approaches in a word discrimination task. Our best approach uses side information in the form of known word pairs to train a Siamese convolutional neural network (CNN): a pair of tied networks that take two speech segments as input and produce their embeddings, trained with a hinge loss that separates same-word pairs and different-word pairs by some margin. A word classifier CNN performs similarly, but requires much stronger supervision. Both types of CNNs yield large improvements over the best previously published results on the word discrimination task.

In settings where only unlabeled speech data is available, speech technology needs to be developed without transcriptions, pronunciation dictionaries, or language modelling text. A similar problem is faced when modeling infant language acquisition. In these cases, categorical linguistic structure needs to be discovered directly from speech audio. We present a novel unsupervised Bayesian model that segments unlabeled speech and clusters the segments into hypothesized word groupings. The result is a complete unsupervised tokenization of the input speech in terms of discovered word types. In our approach, a potential word segment (of arbitrary length) is embedded in a fixed-dimensional acoustic vector space. The model, implemented as a Gibbs sampler, then builds a whole-word acoustic model in this space while jointly performing segmentation. We report word error rates in a small-vocabulary connected digit recognition task by mapping the unsupervised decoded output to ground truth transcriptions. The model achieves around 20% error rate, outperforming a previous HMM-based system by about 10% absolute. Moreover, in contrast to the baseline, our model does not require a pre-specified vocabulary size.

Deep neural networks (DNNs) have become a standard component in supervised ASR, used in both data-driven feature extraction and acoustic modelling. Supervision is typically obtained from a forced alignment that provides phone class targets, requiring transcriptions and pronunciations. We propose a novel unsupervised DNN-based feature extractor that can be trained without these resources in zero-resource settings. Using unsupervised term discovery, we find pairs of isolated word examples of the same unknown type; these provide weak top-down supervision. For each pair, dynamic programming is used to align the feature frames of the two words. Matching frames are presented as input-output pairs to a deep autoencoder (AE) neural network. Using this AE as feature extractor in a word discrimination task, we achieve 64% relative improvement over a previous state-of-the-art system, 57% improvement relative to a bottom-up trained deep AE, and come to within 23% of a supervised system.

Current supervised speech technology relies heavily on tran-scribed speech and pronunciation dictionaries. In settings where unlabelled speech data alone is available, unsupervised methods are required to discover categorical linguistic structure directly from the audio. We present a novel Bayesian model which seg-ments unlabelled input speech into word-like units, resulting in a complete unsupervised transcription of the speech in terms of discovered word types. In our approach, a potential word segment (of arbitrary length) is embedded in a fixed-dimensional space; the model (implemented as a Gibbs sampler) then builds a whole-word acoustic model in this space while jointly doing seg-mentation. We report word error rates in a connected digit recog-nition task by mapping the unsupervised output to ground truth transcriptions. Our model outperforms a previously developed HMM-based system, even when the model is not constrained to discover only the 11 word types present in the data. Index Terms: unsupervised speech processing, word discovery, speech segmentation, unsupervised learning, segmental models 1.

The success of supervised deep neural networks (DNNs) in speech recognition cannot be transferred to zero-resource languages where the requisite transcriptions are unavailable. We investigate unsupervised neural network based methods for learning frame-level representations. Good frame representations eliminate differences in accent, gender, channel characteristics, and other factors to model subword units for within- and across-speaker phonetic discrimination. We enhance the correspondence autoencoder (cAE) and show that it can transform Mel Frequency Cepstral Coefficients (MFCCs) into more effective frame representations given a set of matched word pairs from an unsupervised term discovery (UTD) system. The cAE combines the feature extraction power of autoencoders with the weak supervision signal from UTD pairs to better approximate the extrinsic task’s objective during training. We use the Zero Resource Speech Challenge’s minimal triphone pair ABX discrimination task to evaluate our methods. Optimizing a cAE architecture on English and applying it to a zero-resource language, Xitsonga, we obtain a relative error rate reduction of 35% compared to the original MFCCs. We also show that Xitsonga frame representations extracted from the bottleneck layer of a supervised DNN trained on English can be further enhanced by the cAE, yielding a relative error rate reduction of 39%.

Unsupervised speech processing methods are essential for applications ranging from zero-resource speech technology to modelling child language acquisition. One challenging problem is discovering the word inventory of the language: the lexicon. Lexical clustering is the task of grouping unlabelled acoustic word tokens according to type. We propose a novel lexical clustering model: variable-length word segments are embedded in a fixed-dimensional acoustic space in which clustering is then performed. We evaluate several clustering algorithms and find that the best methods produce clusters with wide variation in sizes, as observed in natural language. The best probabilistic approach is an infinite Gaussian mixture model (IGMM), which automatically chooses the number of clusters. Performance is comparable to that of non-probabilistic Chinese Whispers and average-linkage hierarchical clustering. We conclude that IGMM clustering of fixed-dimensional embeddings holds promise as the lexical clustering component in unsupervised speech processing systems.

In this paper we investigate speech recognition performance of systems employing several accent-specific recognisers in parallel for the simultaneous recognition of multiple accents. We compare these systems with oracle systems, in which test utterances are presented to matching accent-specific recognisers, and with accent-independent systems, in which acoustic and language model training data are pooled. Our investigation is based on Afrikaans (AE), Black (BE) and White (EE) accents of South African English. We find that, when accent is classified on a per-utterance basis, parallel systems outperform oracle systems for the AE+EE accent pair while the opposite is observed for BE+EE. When accent identification is carried out on a per-speaker basis, oracle or better performance is obtained for both accent pairs. Furthermore, parallel systems based on multi-accent acoustic modelling, which allows selective crossaccent sharing of acoustic training data, outperform parallel systems using accent-specific acoustic models. The former also yields better performance than accent-independent recognition, which uses pooled acoustic and language models.