Deep generative models have been widely used for their ability to generate realistic data samples in various areas, such as images, molecules, text, and speech. One major goal of data generation is controllability, namely to generate new data with desired properties. Despite growing interest in the area of controllable generation, significant challenges still remain, including 1) disentangling desired properties with unrelated latent variables, 2) out-of-distribution property control, and 3) objective optimization for out-of-distribution property control. To address these challenges, in this paper, we propose a general framework to enhance VAE-based data generators with property controllability and ensure disentanglement. Our proposed objective can be optimized on both data seen and unseen in the training set. We propose a training procedure to train the objective in a semi-supervised manner by iteratively conducting mutual mappings between the data and properties. The proposed framework is implemented on four VAE-based controllable generators to evaluate its performance on property error, disentanglement, generation quality, and training time. The results indicate that our proposed framework enables more precise control over the properties of generated samples in a short training time, ensuring the disentanglement and keeping the validity of the generated samples.
Large Language models (LLM) have demonstrated the capability to handle a variety of generative tasks. This paper presents the UniAudio system, which, unlike prior task-specific approaches, leverages LLM techniques to generate multiple types of audio (including speech, sounds, music, and singing) with given input conditions. UniAudio 1) first tokenizes all types of target audio along with other condition modalities, 2) concatenates source-target pair as a single sequence, and 3) performs next-token prediction using LLM. Also, a multi-scale Transformer model is proposed to handle the overly long sequences caused by the residual vector quantization based neural codec in tokenization. Training of UniAudio is scaled up to 165K hours of audio and 1B parameters, based on all generative tasks, aiming to obtain sufficient prior knowledge not only in the intrinsic properties of audio but also the inter-relationship between audio and other modalities. Therefore, the trained UniAudio model has the potential to become a foundation model for universal audio generation: it shows strong capability in all trained tasks and can seamlessly support new audio generation tasks after simple fine-tuning. Experiments demonstrate that UniAudio achieves state-of-the-art or at least competitive results on most of the 11 tasks. Demo and code are released at https://github.com/yangdongchao/UniAudio
In this paper, we propose a novel Lip-to-Speech synthesis (L2S) framework, for synthesizing intelligible speech from a silent lip movement video. Specifically, to complement the insufficient supervisory signal of the previous L2S model, we propose to use quantized self-supervised speech representations, named speech units, as an additional prediction target for the L2S model. Therefore, the proposed L2S model is trained to generate multiple targets, mel-spectrogram and speech units. As the speech units are discrete while mel-spectrogram is continuous, the proposed multi-target L2S model can be trained with strong content supervision, without using text-labeled data. Moreover, to accurately convert the synthesized mel-spectrogram into a waveform, we introduce a multi-input vocoder that can generate a clear waveform even from blurry and noisy mel-spectrogram by referring to the speech units. Extensive experimental results confirm the effectiveness of the proposed method in L2S.
Automatic Speech Understanding (ASU) leverages the power of deep learning models for accurate interpretation of human speech, leading to a wide range of speech applications that enrich the human experience. However, training a robust ASU model requires the curation of a large number of speech samples, creating risks for privacy breaches. In this work, we investigate using foundation models to assist privacy-enhancing speech computing. Unlike conventional works focusing primarily on data perturbation or distributed algorithms, our work studies the possibilities of using pre-trained generative models to synthesize speech content as training data with just label guidance. We show that zero-shot learning with training label-guided synthetic speech content remains a challenging task. On the other hand, our results demonstrate that the model trained with synthetic speech samples provides an effective initialization point for low-resource ASU training. This result reveals the potential to enhance privacy by reducing user data collection but using label-guided synthetic speech content.
Background noise considerably reduces the accuracy and reliability of speaker verification (SV) systems. These challenges can be addressed using a speech enhancement system as a front-end module. Recently, diffusion probabilistic models (DPMs) have exhibited remarkable noise-compensation capabilities in the speech enhancement domain. Building on this success, we propose Diff-SV, a noise-robust SV framework that leverages DPM. Diff-SV unifies a DPM-based speech enhancement system with a speaker embedding extractor, and yields a discriminative and noise-tolerable speaker representation through a hierarchical structure. The proposed model was evaluated under both in-domain and out-of-domain noisy conditions using the VoxCeleb1 test set, an external noise source, and the VOiCES corpus. The obtained experimental results demonstrate that Diff-SV achieves state-of-the-art performance, outperforming recently proposed noise-robust SV systems.
In this work, we investigate two popular end-to-end automatic speech recognition (ASR) models, namely Connectionist Temporal Classification (CTC) and RNN-Transducer (RNN-T), for offline recognition of voice search queries, with up to 2B model parameters. The encoders of our models use the neural architecture of Google's universal speech model (USM), with additional funnel pooling layers to significantly reduce the frame rate and speed up training and inference. We perform extensive studies on vocabulary size, time reduction strategy, and its generalization performance on long-form test sets. Despite the speculation that, as the model size increases, CTC can be as good as RNN-T which builds label dependency into the prediction, we observe that a 900M RNN-T clearly outperforms a 1.8B CTC and is more tolerant to severe time reduction, although the WER gap can be largely removed by LM shallow fusion.
We introduce "ivrit.ai", a comprehensive Hebrew speech dataset, addressing the distinct lack of extensive, high-quality resources for advancing Automated Speech Recognition (ASR) technology in Hebrew. With over 3,300 speech hours and a over a thousand diverse speakers, ivrit.ai offers a substantial compilation of Hebrew speech across various contexts. It is delivered in three forms to cater to varying research needs: raw unprocessed audio; data post-Voice Activity Detection, and partially transcribed data. The dataset stands out for its legal accessibility, permitting use at no cost, thereby serving as a crucial resource for researchers, developers, and commercial entities. ivrit.ai opens up numerous applications, offering vast potential to enhance AI capabilities in Hebrew. Future efforts aim to expand ivrit.ai further, thereby advancing Hebrew's standing in AI research and technology.
Speaker verification has been widely explored using speech signals, which has shown significant improvement using deep models. Recently, there has been a surge in exploring faces and voices as they can offer more complementary and comprehensive information than relying only on a single modality of speech signals. Though current methods in the literature on the fusion of faces and voices have shown improvement over that of individual face or voice modalities, the potential of audio-visual fusion is not fully explored for speaker verification. Most of the existing methods based on audio-visual fusion either rely on score-level fusion or simple feature concatenation. In this work, we have explored cross-modal joint attention to fully leverage the inter-modal complementary information and the intra-modal information for speaker verification. Specifically, we estimate the cross-attention weights based on the correlation between the joint feature presentation and that of the individual feature representations in order to effectively capture both intra-modal as well inter-modal relationships among the faces and voices. We have shown that efficiently leveraging the intra- and inter-modal relationships significantly improves the performance of audio-visual fusion for speaker verification. The performance of the proposed approach has been evaluated on the Voxceleb1 dataset. Results show that the proposed approach can significantly outperform the state-of-the-art methods of audio-visual fusion for speaker verification.
In recent years, many interpretability methods have been proposed to help interpret the internal states of Transformer-models, at different levels of precision and complexity. Here, to analyze encoder-decoder Transformers, we propose a simple, new method: DecoderLens. Inspired by the LogitLens (for decoder-only Transformers), this method involves allowing the decoder to cross-attend representations of intermediate encoder layers instead of using the final encoder output, as is normally done in encoder-decoder models. The method thus maps previously uninterpretable vector representations to human-interpretable sequences of words or symbols. We report results from the DecoderLens applied to models trained on question answering, logical reasoning, speech recognition and machine translation. The DecoderLens reveals several specific subtasks that are solved at low or intermediate layers, shedding new light on the information flow inside the encoder component of this important class of models.
The design of automatic speech pronunciation assessment can be categorized into closed and open response scenarios, each with strengths and limitations. A system with the ability to function in both scenarios can cater to diverse learning needs and provide a more precise and holistic assessment of pronunciation skills. In this study, we propose a Multi-task Pronunciation Assessment model called MultiPA. MultiPA provides an alternative to Kaldi-based systems in that it has simpler format requirements and better compatibility with other neural network models. Compared with previous open response systems, MultiPA provides a wider range of evaluations, encompassing assessments at both the sentence and word-level. Our experimental results show that MultiPA achieves comparable performance when working in closed response scenarios and maintains more robust performance when directly used for open responses.