Automatic speech translation is sensitive to speech recognition errors, but in a multilingual scenario, the same content may be available in various languages via simultaneous interpreting, dubbing or subtitling. In this paper, we hypothesize that leveraging multiple sources will improve translation quality if the sources complement one another in terms of correct information they contain. To this end, we first show that on a 10-hour ESIC corpus, the ASR errors in the original English speech and its simultaneous interpreting into German and Czech are mutually independent. We then use two sources, English and German, in a multi-source setting for translation into Czech to establish its robustness to ASR errors. Furthermore, we observe this robustness when translating both noisy sources together in a simultaneous translation setting. Our results show that multi-source neural machine translation has the potential to be useful in a real-time simultaneous translation setting, thereby motivating further investigation in this area.
We present BIG-C (Bemba Image Grounded Conversations), a large multimodal dataset for Bemba. While Bemba is the most populous language of Zambia, it exhibits a dearth of resources which render the development of language technologies or language processing research almost impossible. The dataset is comprised of multi-turn dialogues between Bemba speakers based on images, transcribed and translated into English. There are more than 92,000 utterances/sentences, amounting to more than 180 hours of audio data with corresponding transcriptions and English translations. We also provide baselines on speech recognition (ASR), machine translation (MT) and speech translation (ST) tasks, and sketch out other potential future multimodal uses of our dataset. We hope that by making the dataset available to the research community, this work will foster research and encourage collaboration across the language, speech, and vision communities especially for languages outside the "traditionally" used high-resourced ones. All data and code are publicly available: https://github.com/csikasote/bigc.
Recently, denoising diffusion models have demonstrated remarkable performance among generative models in various domains. However, in the speech domain, the application of diffusion models for synthesizing time-varying audio faces limitations in terms of complexity and controllability, as speech synthesis requires very high-dimensional samples with long-term acoustic features. To alleviate the challenges posed by model complexity in singing voice synthesis, we propose HiddenSinger, a high-quality singing voice synthesis system using a neural audio codec and latent diffusion models. To ensure high-fidelity audio, we introduce an audio autoencoder that can encode audio into an audio codec as a compressed representation and reconstruct the high-fidelity audio from the low-dimensional compressed latent vector. Subsequently, we use the latent diffusion models to sample a latent representation from a musical score. In addition, our proposed model is extended to an unsupervised singing voice learning framework, HiddenSinger-U, to train the model using an unlabeled singing voice dataset. Experimental results demonstrate that our model outperforms previous models in terms of audio quality. Furthermore, the HiddenSinger-U can synthesize high-quality singing voices of speakers trained solely on unlabeled data.
Large documents written in juridical language are difficult to interpret, with long sentences leading to intricate and intertwined relations between the nouns. The present paper frames this problem in the context of recent European security directives. The complexity of their language is here thwarted by automating the extraction of the relevant information, namely of the parts of speech from each clause, through a specific tailoring of Natural Language Processing (NLP) techniques. These contribute, in combination with ontology development principles, to the design of our automated method for the representation of security directives as ontologies. The method is showcased on a practical problem, namely to derive an ontology representing the NIS 2 directive, which is the peak of cybersecurity prescripts at the European level. Although the NLP techniques adopted showed some limitations and had to be complemented by manual analysis, the overall results provide valid support for directive compliance in general and for ontology development in particular.
Current ASR systems are mainly trained and evaluated at the utterance level. Long range cross utterance context can be incorporated. A key task is to derive a suitable compact representation of the most relevant history contexts. In contrast to previous researches based on either LSTM-RNN encoded histories that attenuate the information from longer range contexts, or frame level concatenation of transformer context embeddings, in this paper compact low-dimensional cross utterance contextual features are learned in the Conformer-Transducer Encoder using specially designed attention pooling layers that are applied over efficiently cached preceding utterances history vectors. Experiments on the 1000-hr Gigaspeech corpus demonstrate that the proposed contextualized streaming Conformer-Transducers outperform the baseline using utterance internal context only with statistically significant WER reductions of 0.7% to 0.5% absolute (4.3% to 3.1% relative) on the dev and test data.
Voice synthesis has seen significant improvements in the past decade resulting in highly intelligible voices. Further investigations have resulted in models that can produce variable speech, including conditional emotional expression. The problem lies, however, in a focus on phrase-level modifications and prosodic vocal features. Using the CREMA-D dataset we have trained a GAN conditioned on emotion to generate worth lengths for a given input text. These word lengths are relative to neutral speech and can be provided, through speech synthesis markup language (SSML) to a text-to-speech (TTS) system to generate more expressive speech. Additionally, a generative model is also trained using implicit maximum likelihood estimation (IMLE) and a comparative analysis with GANs is included. We were able to achieve better performances on objective measures for neutral speech, and better time alignment for happy speech when compared to an out-of-box model. However, further investigation of subjective evaluation is required.
DeepFake based digital facial forgery is threatening public media security, especially when lip manipulation has been used in talking face generation, and the difficulty of fake video detection is further improved. By only changing lip shape to match the given speech, the facial features of identity are hard to be discriminated in such fake talking face videos. Together with the lack of attention on audio stream as the prior knowledge, the detection failure of fake talking face videos also becomes inevitable. It's found that the optical flow of the fake talking face video is disordered especially in the lip region while the optical flow of the real video changes regularly, which means the motion feature from optical flow is useful to capture manipulation cues. In this study, a fake talking face detection network (FTFDNet) is proposed by incorporating visual, audio and motion features using an efficient cross-modal fusion (CMF) module. Furthermore, a novel audio-visual attention mechanism (AVAM) is proposed to discover more informative features, which can be seamlessly integrated into any audio-visual CNN architecture by modularization. With the additional AVAM, the proposed FTFDNet is able to achieve a better detection performance than other state-of-the-art DeepFake video detection methods not only on the established fake talking face detection dataset (FTFDD) but also on the DeepFake video detection datasets (DFDC and DF-TIMIT).
Self-supervised learning (SSL) has recently emerged as a promising paradigm for training generalisable models on large-scale data in the fields of vision, text, and speech. Although SSL has been proven effective in speech and audio, its application to music audio has yet to be thoroughly explored. This is primarily due to the distinctive challenges associated with modelling musical knowledge, particularly its tonal and pitched characteristics of music. To address this research gap, we propose an acoustic Music undERstanding model with large-scale self-supervised Training (MERT), which incorporates teacher models to provide pseudo labels in the masked language modelling (MLM) style acoustic pre-training. In our exploration, we identified a superior combination of teacher models, which outperforms conventional speech and audio approaches in terms of performance. This combination includes an acoustic teacher based on Residual Vector Quantization - Variational AutoEncoder (RVQ-VAE) and a musical teacher based on the Constant-Q Transform (CQT). These teachers effectively guide our student model, a BERT-style transformer encoder, to better model music audio. In addition, we introduce an in-batch noise mixture augmentation to enhance the representation robustness. Furthermore, we explore a wide range of settings to overcome the instability in acoustic language model pre-training, which allows our designed paradigm to scale from 95M to 330M parameters. Experimental results indicate that our model can generalise and perform well on 14 music understanding tasks and attains state-of-the-art (SOTA) overall scores. The code and models are online: https://github.com/yizhilll/MERT.
Existing fine-grained intensity regulation methods rely on explicit control through predicted emotion probabilities. However, these high-level semantic probabilities are often inaccurate and unsmooth at the phoneme level, leading to bias in learning. Especially when we attempt to mix multiple emotion intensities for specific phonemes, resulting in markedly reduced controllability and naturalness of the synthesis. To address this issue, we propose the CAScaded Explicit and Implicit coNtrol framework (CASEIN), which leverages accurate disentanglement of emotion manifolds from the reference speech to learn the implicit representation at a lower semantic level. This representation bridges the semantical gap between explicit probabilities and the synthesis model, reducing bias in learning. In experiments, our CASEIN surpasses existing methods in both controllability and naturalness. Notably, we are the first to achieve fine-grained control over the mixed intensity of multiple emotions.
This paper presents a configurable version of Extreme Bandwidth Extension Network (EBEN), a Generative Adversarial Network (GAN) designed to improve audio captured with body-conduction microphones. We show that these microphones significantly reduce environmental noise. However, this insensitivity to ambient noise is at the expense of the bandwidth of the voice signal acquired from the wearer of the devices. The obtained captured signals therefore require the use of signal enhancement techniques to recover the full-bandwidth speech. EBEN leverages a configurable multiband decomposition of the raw captured signal. This decomposition allows the data time domain dimensions to be reduced and the full band signal to be better controlled. The multiband representation of the captured signal is processed through a U-Net-like model, which combines feature and adversarial losses to generate an enhanced speech signal. We also benefit from this original representation in the proposed configurable discriminator architecture. The configurable EBEN approach can achieve state-of-the-art enhancement results on synthetic data with a lightweight generator that allows real-time processing.