Sharon Goldwater

Ramon Sanabria, Hao Tang, Sharon Goldwater

Given the strong results of self-supervised models on various tasks, there have been surprisingly few studies exploring self-supervised representations for acoustic word embeddings (AWE), fixed-dimensional vectors representing variable-length spoken word segments. In this work, we study several pre-trained models and pooling methods for constructing AWEs with self-supervised representations. Owing to the contextualized nature of self-supervised representations, we hypothesize that simple pooling methods, such as averaging, might already be useful for constructing AWEs. When evaluating on a standard word discrimination task, we find that HuBERT representations with mean-pooling rival the state of the art on English AWEs. More surprisingly, despite being trained only on English, HuBERT representations evaluated on Xitsonga, Mandarin, and French consistently outperform the multilingual model XLSR-53 (as well as Wav2Vec 2.0 trained on English).

Burin Naowarat, Hao Tang, Sharon Goldwater

Understanding how neural models represent human-interpretable concepts is challenging. Prior work has explored linear concept subspaces from diverse perspectives, such as probing and concept erasure. We introduce a unified framework to study these subspaces along two axes: \textit{containment}, which tests if a concept is fully represented in a subspace but not outside it, and \textit{disentanglement}, which tests for isolation from other concepts. In experiments on both text and speech models, we first highlight that concept subspaces may not be uniquely determined, and discuss the implications for concept subspace analysis. Then, we compare properties of concept subspaces estimated using five estimators, proposed in different communities. We find that (1) the choice of estimator impacts the containment and disentanglement properties; (2) the state-of-the-art concept erasure method, LEACE, performs well on both testing axes, but still struggles to generalize to unseen data; and (3) in HuBERT speech representations, phone information is both contained and disentangled from speaker information, while speaker information is hard to contain in a compact subspace, despite being disentangled from phones.

Amr Keleg, Sharon Goldwater, Walid Magdy

Transcribed speech and user-generated text in Arabic typically contain a mixture of Modern Standard Arabic (MSA), the standardized language taught in schools, and Dialectal Arabic (DA), used in daily communications. To handle this variation, previous work in Arabic NLP has focused on Dialect Identification (DI) on the sentence or the token level. However, DI treats the task as binary, whereas we argue that Arabic speakers perceive a spectrum of dialectness, which we operationalize at the sentence level as the Arabic Level of Dialectness (ALDi), a continuous linguistic variable. We introduce the AOC-ALDi dataset (derived from the AOC dataset), containing 127,835 sentences (17% from news articles and 83% from user comments on those articles) which are manually labeled with their level of dialectness. We provide a detailed analysis of AOC-ALDi and show that a model trained on it can effectively identify levels of dialectness on a range of other corpora (including dialects and genres not included in AOC-ALDi), providing a more nuanced picture than traditional DI systems. Through case studies, we illustrate how ALDi can reveal Arabic speakers' stylistic choices in different situations, a useful property for sociolinguistic analyses.

Ramon Sanabria, Ondrej Klejch, Hao Tang et al.

Acoustic word embeddings are typically created by training a pooling function using pairs of word-like units. For unsupervised systems, these are mined using k-nearest neighbor (KNN) search, which is slow. Recently, mean-pooled representations from a pre-trained self-supervised English model were suggested as a promising alternative, but their performance on target languages was not fully competitive. Here, we explore improvements to both approaches: we use continued pre-training to adapt the self-supervised model to the target language, and we use a multilingual phone recognizer (MPR) to mine phone n-gram pairs for training the pooling function. Evaluating on four languages, we show that both methods outperform a recent approach on word discrimination. Moreover, the MPR method is orders of magnitude faster than KNN, and is highly data efficient. We also show a small improvement from performing learned pooling on top of the continued pre-trained representations.

Elizabeth Nielsen, Sharon Goldwater, Mark Steedman

Parsing spoken dialogue presents challenges that parsing text does not, including a lack of clear sentence boundaries. We know from previous work that prosody helps in parsing single sentences (Tran et al. 2018), but we want to show the effect of prosody on parsing speech that isn't segmented into sentences. In experiments on the English Switchboard corpus, we find prosody helps our model both with parsing and with accurately identifying sentence boundaries. However, we find that the best-performing parser is not necessarily the parser that produces the best sentence segmentation performance. We suggest that the best parses instead come from modelling sentence boundaries jointly with other constituent boundaries.

Amr Keleg, Walid Magdy, Sharon Goldwater

On annotating multi-dialect Arabic datasets, it is common to randomly assign the samples across a pool of native Arabic speakers. Recent analyses recommended routing dialectal samples to native speakers of their respective dialects to build higher-quality datasets. However, automatically identifying the dialect of samples is hard. Moreover, the pool of annotators who are native speakers of specific Arabic dialects might be scarce. Arabic Level of Dialectness (ALDi) was recently introduced as a quantitative variable that measures how sentences diverge from Standard Arabic. On randomly assigning samples to annotators, we hypothesize that samples of higher ALDi scores are harder to label especially if they are written in dialects that the annotators do not speak. We test this by analyzing the relation between ALDi scores and the annotators' agreement, on 15 public datasets having raw individual sample annotations for various sentence-classification tasks. We find strong evidence supporting our hypothesis for 11 of them. Consequently, we recommend prioritizing routing samples of high ALDi scores to native speakers of each sample's dialect, for which the dialect could be automatically identified at higher accuracies.

Amr Keleg, Sharon Goldwater, Walid Magdy

Arabic has diverse dialects, where one dialect can be substantially different from the others. In the NLP literature, some assumptions about these dialects are widely adopted (e.g., ``Arabic dialects can be grouped into distinguishable regional dialects") and are manifested in different computational tasks such as Arabic Dialect Identification (ADI). However, these assumptions are not quantitatively verified. We identify four of these assumptions and examine them by extending and analyzing a multi-label dataset, where the validity of each sentence in 11 different country-level dialects is manually assessed by speakers of these dialects. Our analysis indicates that the four assumptions oversimplify reality, and some of them are not always accurate. This in turn might be hindering further progress in different Arabic NLP tasks.

Oli Danyi Liu, Hao Tang, Naomi Feldman et al.

Speech perception involves storing and integrating sequentially presented items. Recent work in cognitive neuroscience has identified temporal and contextual characteristics in humans' neural encoding of speech that may facilitate this temporal processing. In this study, we simulated similar analyses with representations extracted from a computational model that was trained on unlabelled speech with the learning objective of predicting upcoming acoustics. Our simulations revealed temporal dynamics similar to those in brain signals, implying that these properties can arise without linguistic knowledge. Another property shared between brains and the model is that the encoding patterns of phonemes support some degree of cross-context generalization. However, we found evidence that the effectiveness of these generalizations depends on the specific contexts, which suggests that this analysis alone is insufficient to support the presence of context-invariant encoding.

Yen Meng, Sharon Goldwater, Hao Tang

Modern neural speech models benefit from having longer context, and many approaches have been proposed to increase the maximum context a model can use. However, few have attempted to measure how much context these models actually use, i.e., the effective context. Here, we propose two approaches to measuring the effective context, and use them to analyze different speech Transformers. For supervised models, we find that the effective context correlates well with the nature of the task, with fundamental frequency tracking, phone classification, and word classification requiring increasing amounts of effective context. For self-supervised models, we find that effective context increases mainly in the early layers, and remains relatively short -- similar to the supervised phone model. Given that these models do not use a long context during prediction, we show that HuBERT can be run in streaming mode without modification to the architecture and without further fine-tuning.

Coleman Haley, Sharon Goldwater, Edoardo Ponti

We propose a grounded approach to meaning in language typology. We treat data from perceptual modalities, such as images, as a language-agnostic representation of meaning. Hence, we can quantify the function--form relationship between images and captions across languages. Inspired by information theory, we define "groundedness", an empirical measure of contextual semantic contentfulness (formulated as a difference in surprisal) which can be computed with multilingual multimodal language models. As a proof of concept, we apply this measure to the typology of word classes. Our measure captures the contentfulness asymmetry between functional (grammatical) and lexical (content) classes across languages, but contradicts the view that functional classes do not convey content. Moreover, we find universal trends in the hierarchy of groundedness (e.g., nouns > adjectives > verbs), and show that our measure partly correlates with psycholinguistic concreteness norms in English. We release a dataset of groundedness scores for 30 languages. Our results suggest that the grounded typology approach can provide quantitative evidence about semantic function in language.

Michele Gubian, Ioana Krehan, Oli Liu et al.

Analyses of self-supervised speech models have begun to reveal where and how they represent different types of information. However, almost all analyses have focused on English. Here, we examine how wav2vec2 models trained on four different languages encode both language-matched and non-matched speech. We use probing classifiers and geometric analyses to examine how phones, lexical tones, and speaker information are represented. We show that for all pretraining and test languages, the subspaces encoding phones, tones, and speakers are largely orthogonal, and that layerwise patterns of probing accuracy are similar, with a relatively small advantage for matched-language phone and tone (but not speaker) probes in the later layers. Our findings suggest that the structure of representations learned by wav2vec2 is largely independent of the speech material used during pretraining.

Mukhtar Mohamed, Oli Danyi Liu, Hao Tang et al.

Self-supervised speech representations can hugely benefit downstream speech technologies, yet the properties that make them useful are still poorly understood. Two candidate properties related to the geometry of the representation space have been hypothesized to correlate well with downstream tasks: (1) the degree of orthogonality between the subspaces spanned by the speaker centroids and phone centroids, and (2) the isotropy of the space, i.e., the degree to which all dimensions are effectively utilized. To study them, we introduce a new measure, Cumulative Residual Variance (CRV), which can be used to assess both properties. Using linear classifiers for speaker and phone ID to probe the representations of six different self-supervised models and two untrained baselines, we ask whether either orthogonality or isotropy correlate with linear probing accuracy. We find that both measures correlate with phonetic probing accuracy, though our results on isotropy are more nuanced.

Oli Liu, Hao Tang, Sharon Goldwater

Self-supervised speech representations are known to encode both speaker and phonetic information, but how they are distributed in the high-dimensional space remains largely unexplored. We hypothesize that they are encoded in orthogonal subspaces, a property that lends itself to simple disentanglement. Applying principal component analysis to representations of two predictive coding models, we identify two subspaces that capture speaker and phonetic variances, and confirm that they are nearly orthogonal. Based on this property, we propose a new speaker normalization method which collapses the subspace that encodes speaker information, without requiring transcriptions. Probing experiments show that our method effectively eliminates speaker information and outperforms a previous baseline in phone discrimination tasks. Moreover, the approach generalizes and can be used to remove information of unseen speakers.

Ida Szubert, Omri Abend, Nathan Schneider et al.

This paper proposes a methodology for constructing such corpora of child directed speech (CDS) paired with sentential logical forms, and uses this method to create two such corpora, in English and Hebrew. The approach enforces a cross-linguistically consistent representation, building on recent advances in dependency representation and semantic parsing. Specifically, the approach involves two steps. First, we annotate the corpora using the Universal Dependencies (UD) scheme for syntactic annotation, which has been developed to apply consistently to a wide variety of domains and typologically diverse languages. Next, we further annotate these data by applying an automatic method for transducing sentential logical forms (LFs) from UD structures. The UD and LF representations have complementary strengths: UD structures are language-neutral and support consistent and reliable annotation by multiple annotators, whereas LFs are neutral as to their syntactic derivation and transparently encode semantic relations. Using this approach, we provide syntactic and semantic annotation for two corpora from CHILDES: Brown's Adam corpus (English; we annotate ~80% of its child-directed utterances), all child-directed utterances from Berman's Hagar corpus (Hebrew). We verify the quality of the UD annotation using an inter-annotator agreement study, and manually evaluate the transduced meaning representations. We then demonstrate the utility of the compiled corpora through (1) a longitudinal corpus study of the prevalence of different syntactic and semantic phenomena in the CDS, and (2) applying an existing computational model of language acquisition to the two corpora and briefly comparing the results across languages.

Ramon Sanabria, Hao Tang, Sharon Goldwater

Word segmentation, the problem of finding word boundaries in speech, is of interest for a range of tasks. Previous papers have suggested that for sequence-to-sequence models trained on tasks such as speech translation or speech recognition, attention can be used to locate and segment the words. We show, however, that even on monolingual data this approach is brittle. In our experiments with different input types, data sizes, and segmentation algorithms, only models trained to predict phones from words succeed in the task. Models trained to predict words from either phones or speech (i.e., the opposite direction needed to generalize to new data), yield much worse results, suggesting that attention-based segmentation is only useful in limited scenarios.

Elizabeth Nielsen, Mark Steedman, Sharon Goldwater

Parsing spoken dialogue poses unique difficulties, including disfluencies and unmarked boundaries between sentence-like units. Previous work has shown that prosody can help with parsing disfluent speech (Tran et al. 2018), but has assumed that the input to the parser is already segmented into sentence-like units (SUs), which isn't true in existing speech applications. We investigate how prosody affects a parser that receives an entire dialogue turn as input (a turn-based model), instead of gold standard pre-segmented SUs (an SU-based model). In experiments on the English Switchboard corpus, we find that when using transcripts alone, the turn-based model has trouble segmenting SUs, leading to worse parse performance than the SU-based model. However, prosody can effectively replace gold standard SU boundaries: with prosody, the turn-based model performs as well as the SU-based model (90.79 vs. 90.65 F1 score, respectively), despite performing two tasks (SU segmentation and parsing) rather than one (parsing alone). Analysis shows that pitch and intensity features are the most important for this corpus, since they allow the model to correctly distinguish an SU boundary from a speech disfluency -- a distinction that the model otherwise struggles to make.

Alexander Robertson, Walid Magdy, Sharon Goldwater

Research in sociology and linguistics shows that people use language not only to express their own identity but to understand the identity of others. Recent work established a connection between expression of identity and emoji usage on social media, through use of emoji skin tone modifiers. Motivated by that finding, this work asks if, as with language, readers are sensitive to such acts of self-expression and use them to understand the identity of authors. In behavioral experiments (n=488), where text and emoji content of social media posts were carefully controlled before being presented to participants, we find in the affirmative -- emoji are a salient signal of author identity. That signal is distinct from, and complementary to, the one encoded in language. Participant groups (based on self-identified ethnicity) showed no differences in how they perceive this signal, except in the case of the default yellow emoji. While both groups associate this with a White identity, the effect was stronger in White participants. Our finding that emoji can index social variables will have experimental applications for researchers but also implications for designers: supposedly ``neutral`` defaults may be more representative of some users than others.

Alexander Robertson, Walid Magdy, Sharon Goldwater

Prior work has shown that Twitter users use skin-toned emoji as an act of self-representation to express their racial/ethnic identity. We test whether this signal of identity can influence readers' perceptions about the content of a post containing that signal. In a large scale (n=944) pre-registered controlled experiment, we manipulate the presence of skin-toned emoji and profile photos in a task where readers rate obscure trivia facts (presented as tweets) as true or false. Using a Bayesian statistical analysis, we find that neither emoji nor profile photo has an effect on how readers rate these facts. This result will be of some comfort to anyone concerned about the manipulation of online users through the crafting of fake profiles.

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.

Aibek Makazhanov, Sharon Goldwater, Adam Lopez

We present LemMED, a character-level encoder-decoder for contextual morphological analysis (combined lemmatization and tagging). LemMED extends and is named after two other attention-based models, namely Lematus, a contextual lemmatizer, and MED, a morphological (re)inflection model. Our approach does not require training separate lemmatization and tagging models, nor does it need additional resources and tools, such as morphological dictionaries or transducers. Moreover, LemMED relies solely on character-level representations and on local context. Although the model can, in principle, account for global context on sentence level, our experiments show that using just a single word of context around each target word is not only more computationally feasible, but yields better results as well. We evaluate LemMED in the framework of the SIMGMORPHON-2019 shared task on combined lemmatization and tagging. In terms of average performance LemMED ranks 5th among 13 systems and is bested only by the submissions that use contextualized embeddings.

Yevgen Matusevych, Thomas Schatz, Herman Kamper et al.

In the first year of life, infants' speech perception becomes attuned to the sounds of their native language. Many accounts of this early phonetic learning exist, but computational models predicting the attunement patterns observed in infants from the speech input they hear have been lacking. A recent study presented the first such model, drawing on algorithms proposed for unsupervised learning from naturalistic speech, and tested it on a single phone contrast. Here we study five such algorithms, selected for their potential cognitive relevance. We simulate phonetic learning with each algorithm and perform tests on three phone contrasts from different languages, comparing the results to infants' discrimination patterns. The five models display varying degrees of agreement with empirical observations, showing that our approach can help decide between candidate mechanisms for early phonetic learning, and providing insight into which aspects of the models are critical for capturing infants' perceptual development.

Herman Kamper, Yevgen Matusevych, Sharon Goldwater

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 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 encodes more phonetic, word duration, language identity and speaker information than the other multilingual models.

Kate McCurdy, Sharon Goldwater, Adam Lopez

Can artificial neural networks learn to represent inflectional morphology and generalize to new words as human speakers do? Kirov and Cotterell (2018) argue that the answer is yes: modern Encoder-Decoder (ED) architectures learn human-like behavior when inflecting English verbs, such as extending the regular past tense form -(e)d to novel words. However, their work does not address the criticism raised by Marcus et al. (1995): that neural models may learn to extend not the regular, but the most frequent class -- and thus fail on tasks like German number inflection, where infrequent suffixes like -s can still be productively generalized. To investigate this question, we first collect a new dataset from German speakers (production and ratings of plural forms for novel nouns) that is designed to avoid sources of information unavailable to the ED model. The speaker data show high variability, and two suffixes evince 'regular' behavior, appearing more often with phonologically atypical inputs. Encoder-decoder models do generalize the most frequently produced plural class, but do not show human-like variability or 'regular' extension of these other plural markers. We conclude that modern neural models may still struggle with minority-class generalization.

Elizabeth Nielsen, Mark Steedman, Sharon Goldwater

Prosody is a rich information source in natural language, serving as a marker for phenomena such as contrast. In order to make this information available to downstream tasks, we need a way to detect prosodic events in speech. We propose a new model for pitch accent detection, inspired by the work of Stehwien et al. (2018), who presented a CNN-based model for this task. Our model makes greater use of context by using full utterances as input and adding an LSTM layer. We find that these innovations lead to an improvement from 87.5% to 88.7% accuracy on pitch accent detection on American English speech in the Boston University Radio News Corpus, a state-of-the-art result. We also find that a simple baseline that just predicts a pitch accent on every content word yields 82.2% accuracy, and we suggest that this is the appropriate baseline for this task. Finally, we conduct ablation tests that show pitch is the most important acoustic feature for this task and this corpus.

Yevgen Matusevych, Herman Kamper, Sharon Goldwater

Recent studies have introduced methods for learning acoustic word embeddings (AWEs)---fixed-size vector representations of words which encode their acoustic features. Despite the widespread use of AWEs in speech processing research, they have only been evaluated quantitatively in their ability to discriminate between whole word tokens. To better understand the applications of AWEs in various downstream tasks and in cognitive modeling, we need to analyze the representation spaces of AWEs. Here we analyze basic properties of AWE spaces learned by a sequence-to-sequence encoder-decoder model in six typologically diverse languages. We first show that these AWEs preserve some information about words' absolute duration and speaker. At the same time, the representation space of these AWEs is organized such that the distance between words' embeddings increases with those words' phonetic dissimilarity. Finally, the AWEs exhibit a word onset bias, similar to patterns reported in various studies on human speech processing and lexical access. We argue this is a promising result and encourage further evaluation of AWEs as a potentially useful tool in cognitive science, which could provide a link between speech processing and lexical memory.

Herman Kamper, Yevgen Matusevych, Sharon Goldwater

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.

Mihaela C. Stoian, Sameer Bansal, Sharon Goldwater

Previous work has shown that for low-resource source languages, automatic speech-to-text translation (AST) can be improved by pretraining an end-to-end model on automatic speech recognition (ASR) data from a high-resource language. However, it is not clear what factors --e.g., language relatedness or size of the pretraining data-- yield the biggest improvements, or whether pretraining can be effectively combined with other methods such as data augmentation. Here, we experiment with pretraining on datasets of varying sizes, including languages related and unrelated to the AST source language. We find that the best predictor of final AST performance is the word error rate of the pretrained ASR model, and that differences in ASR/AST performance correlate with how phonetic information is encoded in the later RNN layers of our model. We also show that pretraining and data augmentation yield complementary benefits for AST.

Sameer Bansal, Herman Kamper, Adam Lopez et al.

Given a large amount of unannotated speech in a low-resource language, can we classify the speech utterances by topic? 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.

Maria Corkery, Yevgen Matusevych, Sharon Goldwater

The cognitive mechanisms needed to account for the English past tense have long been a subject of debate in linguistics and cognitive science. Neural network models were proposed early on, but were shown to have clear flaws. Recently, however, Kirov and Cotterell (2018) showed that modern encoder-decoder (ED) models overcome many of these flaws. They also presented evidence that ED models demonstrate humanlike performance in a nonce-word task. Here, we look more closely at the behaviour of their model in this task. We find that (1) the model exhibits instability across multiple simulations in terms of its correlation with human data, and (2) even when results are aggregated across simulations (treating each simulation as an individual human participant), the fit to the human data is not strong---worse than an older rule-based model. These findings hold up through several alternative training regimes and evaluation measures. Although other neural architectures might do better, we conclude that there is still insufficient evidence to claim that neural nets are a good cognitive model for this task.

Toms Bergmanis, Sharon Goldwater

Lemmatization aims to reduce the sparse data problem by relating the inflected forms of a word to its dictionary form. Using context can help, both for unseen and ambiguous words. Yet most context-sensitive approaches require full lemma-annotated sentences for training, which may be scarce or unavailable in low-resource languages. In addition (as shown here), in a low-resource setting, a lemmatizer can learn more from $n$ labeled examples of distinct words (types) than from $n$ (contiguous) labeled tokens, since the latter contain far fewer distinct types. To combine the efficiency of type-based learning with the benefits of context, we propose a way to train a context-sensitive lemmatizer with little or no labeled corpus data, using inflection tables from the UniMorph project and raw text examples from Wikipedia that provide sentence contexts for the unambiguous UniMorph examples. Despite these being unambiguous examples, the model successfully generalizes from them, leading to improved results (both overall, and especially on unseen words) in comparison to a baseline that does not use context.

Enno Hermann, Herman Kamper, Sharon Goldwater

Subword modeling for zero-resource languages aims to learn low-level representations of speech audio without using transcriptions or other resources from the target language (such as text corpora or pronunciation dictionaries). A good representation should capture phonetic content and abstract away from other types of variability, such as speaker differences and channel noise. Previous work in this area has primarily focused unsupervised learning from target language data only, and has been evaluated only intrinsically. Here we directly compare multiple methods, including some that use only target language speech data and some that use transcribed speech from other (non-target) languages, and we evaluate using two intrinsic measures as well as on a downstream unsupervised word segmentation and clustering task. We find that combining two existing target-language-only methods yields better features than either method alone. Nevertheless, even better results are obtained by extracting target language bottleneck features using a model trained on other languages. Cross-lingual training using just one other language is enough to provide this benefit, but multilingual training helps even more. In addition to these results, which hold across both intrinsic measures and the extrinsic task, we discuss the qualitative differences between the different types of learned features.

Sameer Bansal, Herman Kamper, Karen Livescu et al.

We present a simple approach to improve direct speech-to-text translation (ST) when the source language is low-resource: we pre-train the model on a high-resource automatic speech recognition (ASR) task, and then fine-tune its parameters for ST. We demonstrate that our approach is effective by pre-training on 300 hours of English ASR data to improve Spanish-English ST from 10.8 to 20.2 BLEU when only 20 hours of Spanish-English ST training data are available. Through an ablation study, we find that the pre-trained encoder (acoustic model) accounts for most of the improvement, despite the fact that the shared language in these tasks is the target language text, not the source language audio. Applying this insight, we show that pre-training on ASR helps ST even when the ASR language differs from both source and target ST languages: pre-training on French ASR also improves Spanish-English ST. Finally, we show that the approach improves performance on a true low-resource task: pre-training on a combination of English ASR and French ASR improves Mboshi-French ST, where only 4 hours of data are available, from 3.5 to 7.1 BLEU.

Alexander Robertson, Sharon Goldwater

We highlight several issues in the evaluation of historical text normalization systems that make it hard to tell how well these systems would actually work in practice---i.e., for new datasets or languages; in comparison to more naïve systems; or as a preprocessing step for downstream NLP tools. We illustrate these issues and exemplify our proposed evaluation practices by comparing two neural models against a naïve baseline system. We show that the neural models generalize well to unseen words in tests on five languages; nevertheless, they provide no clear benefit over the naïve baseline for downstream POS tagging of an English historical collection. We conclude that future work should include more rigorous evaluation, including both intrinsic and extrinsic measures where possible.

Sameer Bansal, Herman Kamper, Karen Livescu et al.

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.

Enno Hermann, Sharon Goldwater

How can we effectively develop speech technology for languages where no transcribed data is available? Many existing approaches use no annotated resources at all, yet it makes sense to leverage information from large annotated corpora in other languages, for example in the form of multilingual bottleneck features (BNFs) obtained from a supervised speech recognition system. In this work, we evaluate the benefits of BNFs for subword modeling (feature extraction) in six unseen languages on a word discrimination task. First we establish a strong unsupervised baseline by combining two existing methods: vocal tract length normalisation (VTLN) and the correspondence autoencoder (cAE). We then show that BNFs trained on a single language already beat this baseline; including up to 10 languages results in additional improvements which cannot be matched by just adding more data from a single language. Finally, we show that the cAE can improve further on the BNFs if high-quality same-word pairs are available.

Herman Kamper, Karen Livescu, Sharon Goldwater

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.

Sameer Bansal, Herman Kamper, Adam Lopez et al.

We explore the problem of translating speech to text in low-resource scenarios where neither automatic speech recognition (ASR) nor machine translation (MT) are available, but we have training data in the form of audio paired with text translations. We present the first system for this problem applied to a realistic multi-speaker dataset, the CALLHOME Spanish-English speech translation corpus. Our approach uses unsupervised term discovery (UTD) to cluster repeated patterns in the audio, creating a pseudotext, which we pair with translations to create a parallel text and train a simple bag-of-words MT model. We identify the challenges faced by the system, finding that the difficulty of cross-speaker UTD results in low recall, but that our system is still able to correctly translate some content words in test data.

Sameer Bansal, Herman Kamper, Sharon Goldwater et al.

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.

Herman Kamper, Aren Jansen, Sharon Goldwater

Zero-resource speech technology is a growing research area that aims to develop methods for speech processing in the absence of transcriptions, lexicons, or language modelling text. Early term discovery systems focused on identifying isolated recurring patterns in a corpus, while more recent full-coverage systems attempt to completely segment and cluster the audio into word-like units---effectively performing unsupervised speech recognition. This article presents the first attempt we are aware of to apply such a system to large-vocabulary multi-speaker data. Our system uses a Bayesian modelling framework with segmental word representations: each word segment is represented as a fixed-dimensional acoustic embedding obtained by mapping the sequence of feature frames to a single embedding vector. We compare our system on English and Xitsonga datasets to state-of-the-art baselines, using a variety of measures including word error rate (obtained by mapping the unsupervised output to ground truth transcriptions). Very high word error rates are reported---in the order of 70--80% for speaker-dependent and 80--95% for speaker-independent systems---highlighting the difficulty of this task. Nevertheless, in terms of cluster quality and word segmentation metrics, we show that by imposing a consistent top-down segmentation while also using bottom-up knowledge from detected syllable boundaries, both single-speaker and multi-speaker versions of our system outperform a purely bottom-up single-speaker syllable-based approach. We also show that the discovered clusters can be made less speaker- and gender-specific by using an unsupervised autoencoder-like feature extractor to learn better frame-level features (prior to embedding). Our system's discovered clusters are still less pure than those of unsupervised term discovery systems, but provide far greater coverage.

Herman Kamper, Aren Jansen, Sharon Goldwater

In settings where only unlabelled speech data is available, speech technology needs to be developed without transcriptions, pronunciation dictionaries, or language modelling text. A similar problem is faced when modelling 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 unlabelled 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.