In this paper, we present MuLanTTS, the Microsoft end-to-end neural text-to-speech (TTS) system designed for the Blizzard Challenge 2023. About 50 hours of audiobook corpus for French TTS as hub task and another 2 hours of speaker adaptation as spoke task are released to build synthesized voices for different test purposes including sentences, paragraphs, homographs, lists, etc. Building upon DelightfulTTS, we adopt contextual and emotion encoders to adapt the audiobook data to enrich beyond sentences for long-form prosody and dialogue expressiveness. Regarding the recording quality, we also apply denoise algorithms and long audio processing for both corpora. For the hub task, only the 50-hour single speaker data is used for building the TTS system, while for the spoke task, a multi-speaker source model is used for target speaker fine tuning. MuLanTTS achieves mean scores of quality assessment 4.3 and 4.5 in the respective tasks, statistically comparable with natural speech while keeping good similarity according to similarity assessment. The excellent and similarity in this year's new and dense statistical evaluation show the effectiveness of our proposed system in both tasks.
Large Language Models (LLMs) have demonstrated impressive performance on Natural Language Processing (NLP) tasks, such as Question Answering, Summarization, and Classification. The use of LLMs as evaluators, that can rank or score the output of other models (usually LLMs) has become increasingly popular, due to the limitations of current evaluation techniques including the lack of appropriate benchmarks, metrics, cost, and access to human annotators. While LLMs are capable of handling approximately 100 languages, the majority of languages beyond the top 20 lack systematic evaluation across various tasks, metrics, and benchmarks. This creates an urgent need to scale up multilingual evaluation to ensure a precise understanding of LLM performance across diverse languages. LLM-based evaluators seem like the perfect solution to this problem, as they do not require human annotators, human-created references, or benchmarks and can theoretically be used to evaluate any language covered by the LLM. In this paper, we investigate whether LLM-based evaluators can help scale up multilingual evaluation. Specifically, we calibrate LLM-based evaluation against 20k human judgments of five metrics across three text-generation tasks in eight languages. Our findings indicate that LLM-based evaluators may exhibit bias towards higher scores and should be used with caution and should always be calibrated with a dataset of native speaker judgments, particularly in low-resource and non-Latin script languages.
In this paper, we propose PhantomSound, a query-efficient black-box attack toward voice assistants. Existing black-box adversarial attacks on voice assistants either apply substitution models or leverage the intermediate model output to estimate the gradients for crafting adversarial audio samples. However, these attack approaches require a significant amount of queries with a lengthy training stage. PhantomSound leverages the decision-based attack to produce effective adversarial audios, and reduces the number of queries by optimizing the gradient estimation. In the experiments, we perform our attack against 4 different speech-to-text APIs under 3 real-world scenarios to demonstrate the real-time attack impact. The results show that PhantomSound is practical and robust in attacking 5 popular commercial voice controllable devices over the air, and is able to bypass 3 liveness detection mechanisms with >95% success rate. The benchmark result shows that PhantomSound can generate adversarial examples and launch the attack in a few minutes. We significantly enhance the query efficiency and reduce the cost of a successful untargeted and targeted adversarial attack by 93.1% and 65.5% compared with the state-of-the-art black-box attacks, using merely ~300 queries (~5 minutes) and ~1,500 queries (~25 minutes), respectively.
Previous studies have typically assumed that large language models are unable to accurately perform arithmetic operations, particularly multiplication of >8 digits, and operations involving decimals and fractions, without the use of calculator tools. This paper aims to challenge this misconception. With sufficient training data, a 2 billion-parameter language model can accurately perform multi-digit arithmetic operations with almost 100% accuracy without data leakage, significantly surpassing GPT-4 (whose multi-digit multiplication accuracy is only 4.3%). We also demonstrate that our MathGLM, fine-tuned from GLM-10B on a dataset with additional multi-step arithmetic operations and math problems described in text, achieves similar performance to GPT-4 on a 5,000-samples Chinese math problem test set.
Text recognition in the wild is a long-standing problem in computer vision. Driven by end-to-end deep learning, recent studies suggest vision and language processing are effective for scene text recognition. Yet, solving edit errors such as add, delete, or replace is still the main challenge for existing approaches. In fact, the content of the text and its audio are naturally corresponding to each other, i.e., a single character error may result in a clear different pronunciation. In this paper, we propose the AudioOCR, a simple yet effective probabilistic audio decoder for mel spectrogram sequence prediction to guide the scene text recognition, which only participates in the training phase and brings no extra cost during the inference stage. The underlying principle of AudioOCR can be easily applied to the existing approaches. Experiments using 7 previous scene text recognition methods on 12 existing regular, irregular, and occluded benchmarks demonstrate our proposed method can bring consistent improvement. More importantly, through our experimentation, we show that AudioOCR possesses a generalizability that extends to more challenging scenarios, including recognizing non-English text, out-of-vocabulary words, and text with various accents. Code will be available at https://github.com/wenwenyu/AudioOCR.
This paper addresses the problem of selecting of a set of texts for annotation in text classification using retrieval methods when there are limits on the number of annotations due to constraints on human resources. An additional challenge addressed is dealing with binary categories that have a small number of positive instances, reflecting severe class imbalance. In our situation, where annotation occurs over a long time period, the selection of texts to be annotated can be made in batches, with previous annotations guiding the choice of the next set. To address these challenges, the paper proposes leveraging SHAP to construct a quality set of queries for Elasticsearch and semantic search, to try to identify optimal sets of texts for annotation that will help with class imbalance. The approach is tested on sets of cue texts describing possible future events, constructed by participants involved in studies aimed to help with the management of obesity and diabetes. We introduce an effective method for selecting a small set of texts for annotation and building high-quality classifiers. We integrate vector search, semantic search, and machine learning classifiers to yield a good solution. Our experiments demonstrate improved F1 scores for the minority classes in binary classification.
While Large Language Models (LLMs) have demonstrated commendable performance across a myriad of domains and tasks, existing LLMs still exhibit a palpable deficit in handling multimodal functionalities, especially for the Spoken Question Answering (SQA) task which necessitates precise alignment and deep interaction between speech and text features. To address the SQA challenge on LLMs, we initially curated the free-form and open-ended LibriSQA dataset from Librispeech, comprising Part I with natural conversational formats and Part II encompassing multiple-choice questions followed by answers and analytical segments. Both parts collectively include 107k SQA pairs that cover various topics. Given the evident paucity of existing speech-text LLMs, we propose a lightweight, end-to-end framework to execute the SQA task on the LibriSQA, witnessing significant results. By reforming ASR into the SQA format, we further substantiate our framework's capability in handling ASR tasks. Our empirical findings bolster the LLMs' aptitude for aligning and comprehending multimodal information, paving the way for the development of universal multimodal LLMs. The dataset and demo can be found at https://github.com/ZihanZhaoSJTU/LibriSQA.
This paper aims to re-assess scene text recognition (STR) from a data-oriented perspective. We begin by revisiting the six commonly used benchmarks in STR and observe a trend of performance saturation, whereby only 2.91% of the benchmark images cannot be accurately recognized by an ensemble of 13 representative models. While these results are impressive and suggest that STR could be considered solved, however, we argue that this is primarily due to the less challenging nature of the common benchmarks, thus concealing the underlying issues that STR faces. To this end, we consolidate a large-scale real STR dataset, namely Union14M, which comprises 4 million labeled images and 10 million unlabeled images, to assess the performance of STR models in more complex real-world scenarios. Our experiments demonstrate that the 13 models can only achieve an average accuracy of 66.53% on the 4 million labeled images, indicating that STR still faces numerous challenges in the real world. By analyzing the error patterns of the 13 models, we identify seven open challenges in STR and develop a challenge-driven benchmark consisting of eight distinct subsets to facilitate further progress in the field. Our exploration demonstrates that STR is far from being solved and leveraging data may be a promising solution. In this regard, we find that utilizing the 10 million unlabeled images through self-supervised pre-training can significantly improve the robustness of STR model in real-world scenarios and leads to state-of-the-art performance.
Neural networks have been able to generate high-quality single-sentence speech with substantial expressiveness. However, it remains a challenge concerning paragraph-level speech synthesis due to the need for coherent acoustic features while delivering fluctuating speech styles. Meanwhile, training these models directly on over-length speech leads to a deterioration in the quality of synthesis speech. To address these problems, we propose a high-quality and expressive paragraph speech synthesis system with a multi-step variational autoencoder. Specifically, we employ multi-step latent variables to capture speech information at different grammatical levels before utilizing these features in parallel to generate speech waveform. We also propose a three-step training method to improve the decoupling ability. Our model was trained on a single-speaker French audiobook corpus released at Blizzard Challenge 2023. Experimental results underscore the significant superiority of our system over baseline models.
Musicians and fans often produce lyric videos, a form of music videos that showcase the song's lyrics, for their favorite songs. However, making such videos can be challenging and time-consuming as the lyrics need to be added in synchrony and visual harmony with the video. Informed by prior work and close examination of existing lyric videos, we propose a set of design guidelines to help creators make such videos. Our guidelines ensure the readability of the lyric text while maintaining a unified focus of attention. We instantiate these guidelines in a fully automated pipeline that converts an input music video into a lyric video. We demonstrate the robustness of our pipeline by generating lyric videos from a diverse range of input sources. A user study shows that lyric videos generated by our pipeline are effective in maintaining text readability and unifying the focus of attention.