Code-switching (CS) refers to the switching of languages within a speech signal and results in language confusion for automatic speech recognition (ASR). To address language confusion, we propose the language alignment loss that performs frame-level language identification using pseudo language labels learned from the ASR decoder. This eliminates the need for frame-level language annotations. To further tackle the complex token alternatives for language modeling in bilingual scenarios, we propose to employ large language models via a generative error correction method. A linguistic hint that incorporates language information (derived from the proposed language alignment loss and decoded hypotheses) is introduced to guide the prompting of large language models. The proposed methods are evaluated on the SEAME dataset and data from the ASRU 2019 Mandarin-English code-switching speech recognition challenge. The incorporation of the proposed language alignment loss demonstrates a higher CS-ASR performance with only a negligible increase in the number of parameters on both datasets compared to the baseline model. This work also highlights the efficacy of language alignment loss in balancing primary-language-dominant bilingual data during training, with an 8.6% relative improvement on the ASRU dataset compared to the baseline model. Performance evaluation using large language models reveals the advantage of the linguistic hint by achieving 14.1% and 5.5% relative improvement on test sets of the ASRU and SEAME datasets, respectively.
Conversation is the subject of increasing interest in the social, cognitive, and computational sciences. And yet, as conversational datasets continue to increase in size and complexity, researchers lack scalable methods to segment speech-to-text transcripts into conversational turns--the basic building blocks of social interaction. We introduce "NaturalTurn," a turn segmentation algorithm designed to accurately capture the dynamics of naturalistic exchange. NaturalTurn operates by distinguishing speakers' primary conversational turns from listeners' secondary utterances, such as backchannels, brief interjections, and other forms of parallel speech that characterize conversation. Using data from a large conversation corpus, we show how NaturalTurn-derived transcripts demonstrate favorable statistical and inferential characteristics compared to transcripts derived from existing methods. The NaturalTurn algorithm represents an improvement in machine-generated transcript processing methods, or "turn models" that will enable researchers to associate turn-taking dynamics with the broader outcomes that result from social interaction, a central goal of conversation science.
Non-verbal signals in speech are encoded by prosody and carry information that ranges from conversation action to attitude and emotion. Despite its importance, the principles that govern prosodic structure are not yet adequately understood. This paper offers an analytical schema and a technological proof-of-concept for the categorization of prosodic signals and their association with meaning. The schema interprets surface-representations of multi-layered prosodic events. As a first step towards implementation, we present a classification process that disentangles prosodic phenomena of three orders. It relies on fine-tuning a pre-trained speech recognition model, enabling the simultaneous multi-class/multi-label detection. It generalizes over a large variety of spontaneous data, performing on a par with, or superior to, human annotation. In addition to a standardized formalization of prosody, disentangling prosodic patterns can direct a theory of communication and speech organization. A welcome by-product is an interpretation of prosody that will enhance speech- and language-related technologies.
This paper provides a computational analysis of poetry reading audio signals at a large scale to unveil the musicality within professionally-read poems. Although the acoustic characteristics of other types of spoken language have been extensively studied, most of the literature is limited to narrative speech or singing voice, discussing how different they are from each other. In this work, we develop signal processing methods, which are tailored to capture the unique acoustic characteristics of poetry reading based on their silence patterns, temporal variations of local pitch, and beat stability. Our large-scale statistical analyses on three big corpora, each of which consists of narration (LibriSpeech), singing voice (Intonation), and poetry reading (from The Poetry Foundation), discover that poetry reading does share some musical characteristics with singing voice, although it may also resemble narrative speech.
The current dominant approach for neural speech enhancement is based on supervised learning by using simulated training data. The trained models, however, often exhibit limited generalizability to real-recorded data. To address this, we investigate training models directly on real target-domain data, and propose two algorithms, mixture-to-mixture (M2M) training and a co-learning algorithm that improves M2M with the help of supervised algorithms. When paired close-talk and far-field mixtures are available for training, M2M realizes speech enhancement by training a deep neural network (DNN) to produce speech and noise estimates in a way such that they can be linearly filtered to reconstruct the close-talk and far-field mixtures. This way, the DNN can be trained directly on real mixtures, and can leverage close-talk mixtures as a weak supervision to enhance far-field mixtures. To improve M2M, we combine it with supervised approaches to co-train the DNN, where mini-batches of real close-talk and far-field mixture pairs and mini-batches of simulated mixture and clean speech pairs are alternately fed to the DNN, and the loss functions are respectively (a) the mixture reconstruction loss on the real close-talk and far-field mixtures and (b) the regular enhancement loss on the simulated clean speech and noise. We find that, this way, the DNN can learn from real and simulated data to achieve better generalization to real data. We name this algorithm SuperME, $\underline{super}$vised and $\underline{m}$ixture-to-mixtur$\underline{e}$ co-learning. Evaluation results on the CHiME-4 dataset show its effectiveness and potential.
The recent development of decentralised and interoperable social networks (such as the "fediverse") creates new challenges for content moderators. This is because millions of posts generated on one server can easily "spread" to another, even if the recipient server has very different moderation policies. An obvious solution would be to leverage moderation tools to automatically tag (and filter) posts that contravene moderation policies, e.g. related to toxic speech. Recent work has exploited the conversational context of a post to improve this automatic tagging, e.g. using the replies to a post to help classify if it contains toxic speech. This has shown particular potential in environments with large training sets that contain complete conversations. This, however, creates challenges in a decentralised context, as a single conversation may be fragmented across multiple servers. Thus, each server only has a partial view of an entire conversation because conversations are often federated across servers in a non-synchronized fashion. To address this, we propose a decentralised conversation-aware content moderation approach suitable for the fediverse. Our approach employs a graph deep learning model (GraphNLI) trained locally on each server. The model exploits local data to train a model that combines post and conversational information captured through random walks to detect toxicity. We evaluate our approach with data from Pleroma, a major decentralised and interoperable micro-blogging network containing 2 million conversations. Our model effectively detects toxicity on larger instances, exclusively trained using their local post information (0.8837 macro-F1). Our approach has considerable scope to improve moderation in decentralised and interoperable social networks such as Pleroma or Mastodon.
In this work, we extend our previously proposed offline SpatialNet for long-term streaming multichannel speech enhancement in both static and moving speaker scenarios. SpatialNet exploits spatial information, such as the spatial/steering direction of speech, for discriminating between target speech and interferences, and achieved outstanding performance. The core of SpatialNet is a narrow-band self-attention module used for learning the temporal dynamic of spatial vectors. Towards long-term streaming speech enhancement, we propose to replace the offline self-attention network with online networks that have linear inference complexity w.r.t signal length and meanwhile maintain the capability of learning long-term information. Three variants are developed based on (i) masked self-attention, (ii) Retention, a self-attention variant with linear inference complexity, and (iii) Mamba, a structured-state-space-based RNN-like network. Moreover, we investigate the length extrapolation ability of different networks, namely test on signals that are much longer than training signals, and propose a short-signal training plus long-signal fine-tuning strategy, which largely improves the length extrapolation ability of the networks within limited training time. Overall, the proposed online SpatialNet achieves outstanding speech enhancement performance for long audio streams, and for both static and moving speakers. The proposed method will be open-sourced in https://github.com/Audio-WestlakeU/NBSS.
Research in speech technologies and comparative linguistics depends on access to diverse and accessible speech data. The UCLA Phonetics Lab Archive is one of the earliest multilingual speech corpora, with long-form audio recordings and phonetic transcriptions for 314 languages (Ladefoged et al., 2009). Recently, 95 of these languages were time-aligned with word-level phonetic transcriptions (Li et al., 2021). Here we present VoxAngeles, a corpus of audited phonetic transcriptions and phone-level alignments of the UCLA Phonetics Lab Archive, which uses the 95-language CMU re-release as our starting point. VoxAngeles also includes word- and phone-level segmentations from the original UCLA corpus, as well as phonetic measurements of word and phone durations, vowel formants, and vowel f0. This corpus enhances the usability of the original data, particularly for quantitative phonetic typology, as demonstrated through a case study of vowel intrinsic f0. We also discuss the utility of the VoxAngeles corpus for general research and pedagogy in crosslinguistic phonetics, as well as for low-resource and multilingual speech technologies. VoxAngeles is free to download and use under a CC-BY-NC 4.0 license.
As Automatic Speech Recognition (ASR) models become ever more pervasive, it is important to ensure that they make reliable predictions under corruptions present in the physical and digital world. We propose Speech Robust Bench (SRB), a comprehensive benchmark for evaluating the robustness of ASR models to diverse corruptions. SRB is composed of 69 input perturbations which are intended to simulate various corruptions that ASR models may encounter in the physical and digital world. We use SRB to evaluate the robustness of several state-of-the-art ASR models and observe that model size and certain modeling choices such as discrete representations, and self-training appear to be conducive to robustness. We extend this analysis to measure the robustness of ASR models on data from various demographic subgroups, namely English and Spanish speakers, and males and females, and observed noticeable disparities in the model's robustness across subgroups. We believe that SRB will facilitate future research towards robust ASR models, by making it easier to conduct comprehensive and comparable robustness evaluations.
A generative adversarial network (GAN)-based vocoder trained with an adversarial discriminator is commonly used for speech synthesis because of its fast, lightweight, and high-quality characteristics. However, this data-driven model requires a large amount of training data incurring high data-collection costs. This fact motivates us to train a GAN-based vocoder on limited data. A promising solution is to augment the training data to avoid overfitting. However, a standard discriminator is unconditional and insensitive to distributional changes caused by data augmentation. Thus, augmented speech (which can be extraordinary) may be considered real speech. To address this issue, we propose an augmentation-conditional discriminator (AugCondD) that receives the augmentation state as input in addition to speech, thereby assessing the input speech according to the augmentation state, without inhibiting the learning of the original non-augmented distribution. Experimental results indicate that AugCondD improves speech quality under limited data conditions while achieving comparable speech quality under sufficient data conditions. Audio samples are available at https://www.kecl.ntt.co.jp/people/kaneko.takuhiro/projects/augcondd/.