Ge Zhu

Yinghao Aaron Li, Xilin Jiang, Jordan Darefsky et al.

The rapid advancement of large language models (LLMs) has significantly propelled the development of text-based chatbots, demonstrating their capability to engage in coherent and contextually relevant dialogues. However, extending these advancements to enable end-to-end speech-to-speech conversation bots remains a formidable challenge, primarily due to the extensive dataset and computational resources required. The conventional approach of cascading automatic speech recognition (ASR), LLM, and text-to-speech (TTS) models in a pipeline, while effective, suffers from unnatural prosody because it lacks direct interactions between the input audio and its transcribed text and the output audio. These systems are also limited by their inherent latency from the ASR process for real-time applications. This paper introduces Style-Talker, an innovative framework that fine-tunes an audio LLM alongside a style-based TTS model for fast spoken dialog generation. Style-Talker takes user input audio and uses transcribed chat history and speech styles to generate both the speaking style and text for the response. Subsequently, the TTS model synthesizes the speech, which is then played back to the user. While the response speech is being played, the input speech undergoes ASR processing to extract the transcription and speaking style, serving as the context for the ensuing dialogue turn. This novel pipeline accelerates the traditional cascade ASR-LLM-TTS systems while integrating rich paralinguistic information from input speech. Our experimental results show that Style-Talker significantly outperforms the conventional cascade and speech-to-speech baselines in terms of both dialogue naturalness and coherence while being more than 50% faster.

Ge Zhu, Juan-Pablo Caceres, Justin Salamon

Filler words such as `uh' or `um' are sounds or words people use to signal they are pausing to think. Finding and removing filler words from recordings is a common and tedious task in media editing. Automatically detecting and classifying filler words could greatly aid in this task, but few studies have been published on this problem to date. A key reason is the absence of a dataset with annotated filler words for model training and evaluation. In this work, we present a novel speech dataset, PodcastFillers, with 35K annotated filler words and 50K annotations of other sounds that commonly occur in podcasts such as breaths, laughter, and word repetitions. We propose a pipeline that leverages VAD and ASR to detect filler candidates and a classifier to distinguish between filler word types. We evaluate our proposed pipeline on PodcastFillers, compare to several baselines, and present a detailed ablation study. In particular, we evaluate the importance of using ASR and how it compares to a transcription-free approach resembling keyword spotting. We show that our pipeline obtains state-of-the-art results, and that leveraging ASR strongly outperforms a keyword spotting approach. We make PodcastFillers publicly available, in the hope that our work serves as a benchmark for future research.

Ge Zhu, Yujia Yan, Juan-Pablo Caceres et al.

Non-linguistic filler words, such as "uh" or "um", are prevalent in spontaneous speech and serve as indicators for expressing hesitation or uncertainty. Previous works for detecting certain non-linguistic filler words are highly dependent on transcriptions from a well-established commercial automatic speech recognition (ASR) system. However, certain ASR systems are not universally accessible from many aspects, e.g., budget, target languages, and computational power. In this work, we investigate filler word detection system that does not depend on ASR systems. We show that, by using the structured state space sequence model (S4) and neural semi-Markov conditional random fields (semi-CRFs), we achieve an absolute F1 improvement of 6.4% (segment level) and 3.1% (event level) on the PodcastFillers dataset. We also conduct a qualitative analysis on the detected results to analyze the limitations of our proposed system.

Shih-Lun Wu, Ge Zhu, Juan-Pablo Caceres et al.

Music stem generation, the task of producing musically-synchronized and isolated instrument audio clips, offers the potential of greater user control and better alignment with musician workflows compared to conventional text-to-music models. Existing stem generation approaches, however, either rely on fixed architectures that output a predefined set of stems in parallel, or generate only one stem at a time, resulting in slow inference despite flexibility in stem combination. We propose Stemphonic, a diffusion-/flow-based framework that overcomes this trade-off and generates a variable set of synchronized stems in one inference pass. During training, we treat each stem as a batch element, group synchronized stems in a batch, and apply a shared noise latent to each group. At inference-time, we use a shared initial noise latent and stem-specific text inputs to generate synchronized multi-stem outputs in one pass. We further expand our approach to enable one-pass conditional multi-stem generation and stem-wise activity controls to empower users to iteratively generate and orchestrate the temporal layering of a mix. We benchmark our results on multiple open-source stem evaluation sets and show that Stemphonic produces higher-quality outputs while accelerating the full mix generation process by 25 to 50%. Demos at: https://stemphonic-demo.vercel.app.

Ge Zhu, Jinbao Li, Yahong Guo

Current methods aggregate multi-level features or introduce edge and skeleton to get more refined saliency maps. However, little attention is paid to how to obtain the complete salient object in cluttered background, where the targets are usually similar in color and texture to the background. To handle this complex scene, we propose a sharp eyes network (SENet) that first seperates the object from scene, and then finely segments it, which is in line with human visual characteristics, i.e., to look first and then focus. Different from previous methods which directly integrate edge or skeleton to supplement the defects of objects, the proposed method aims to utilize the expanded objects to guide the network obtain complete prediction. Specifically, SENet mainly consists of target separation (TS) brach and object segmentation (OS) branch trained by minimizing a new hierarchical difference aware (HDA) loss. In the TS branch, we construct a fractal structure to produce saliency features with expanded boundary via the supervision of expanded ground truth, which can enlarge the detail difference between foreground and background. In the OS branch, we first aggregate multi-level features to adaptively select complementary components, and then feed the saliency features with expanded boundary into aggregated features to guide the network obtain complete prediction. Moreover, we propose the HDA loss to further improve the structural integrity and local details of the salient objects, which assigns weight to each pixel according to its distance from the boundary hierarchically. Hard pixels with similar appearance in border region will be given more attention hierarchically to emphasize their importance in completeness prediction. Comprehensive experimental results on five datasets demonstrate that the proposed approach outperforms the state-of-the-art methods both quantitatively and qualitatively.

Xuexiang Niu, Jinping Tang, Lei Wang et al.

Learning from feedback has been shown to enhance the alignment between text prompts and images in text-to-image diffusion models. However, due to the lack of focus in feedback content, especially regarding the object type and quantity, these techniques struggle to accurately match text and images when faced with specified prompts. To address this issue, we propose an efficient fine-turning method with specific reward objectives, including three stages. First, generated images from diffusion model are detected to obtain the object categories and quantities. Meanwhile, the confidence of category and quantity can be derived from the detection results and given prompts. Next, we define a novel matching score, based on above confidence, to measure text-image alignment. It can guide the model for feedback learning in the form of a reward function. Finally, we fine-tune the diffusion model by backpropagation the reward function gradients to generate semantically related images. Different from previous feedbacks that focus more on overall matching, we place more emphasis on the accuracy of entity categories and quantities. Besides, we construct a text-to-image dataset for studying the compositional generation, including 1.7 K pairs of text-image with diverse combinations of entities and quantities. Experimental results on this benchmark show that our model outperforms other SOTA methods in both alignment and fidelity. In addition, our model can also serve as a metric for evaluating text-image alignment in other models. All code and dataset are available at https://github.com/kingniu0329/Visions.

You Zhang, Ge Zhu, Zhiyao Duan

The performance of automatic speaker verification (ASV) systems could be degraded by voice spoofing attacks. Most existing works aimed to develop standalone spoofing countermeasure (CM) systems. Relatively little work targeted at developing an integrated spoofing aware speaker verification (SASV) system. In the recent SASV challenge, the organizers encourage the development of such integration by releasing official protocols and baselines. In this paper, we build a probabilistic framework for fusing the ASV and CM subsystem scores. We further propose fusion strategies for direct inference and fine-tuning to predict the SASV score based on the framework. Surprisingly, these strategies significantly improve the SASV equal error rate (EER) from 19.31% of the baseline to 1.53% on the official evaluation trials of the SASV challenge. We verify the effectiveness of our proposed components through ablation studies and provide insights with score distribution analysis.

Ge Zhu, Frank Cwitkowitz, Zhiyao Duan

In this paper, we conduct a cross-dataset study on parametric and non-parametric raw-waveform based speaker embeddings through speaker verification experiments. In general, we observe a more significant performance degradation of these raw-waveform systems compared to spectral based systems. We then propose two strategies to improve the performance of raw-waveform based systems on cross-dataset tests. The first strategy is to change the real-valued filters into analytic filters to ensure shift-invariance. The second strategy is to apply variational dropout to non-parametric filters to prevent them from overfitting irrelevant nuance features.

Xinhui Chen, You Zhang, Ge Zhu et al.

In this paper, we present UR-AIR system submission to the logical access (LA) and the speech deepfake (DF) tracks of the ASVspoof 2021 Challenge. The LA and DF tasks focus on synthetic speech detection (SSD), i.e. detecting text-to-speech and voice conversion as spoofing attacks. Different from previous ASVspoof challenges, the LA task this year presents codec and transmission channel variability, while the new task DF presents general audio compression. Built upon our previous research work on improving the robustness of the SSD systems to channel effects, we propose a channel-robust synthetic speech detection system for the challenge. To mitigate the channel variability issue, we use an acoustic simulator to apply transmission codec, compression codec, and convolutional impulse responses to augmenting the original datasets. For the neural network backbone, we propose to use Emphasized Channel Attention, Propagation and Aggregation Time Delay Neural Networks (ECAPA-TDNN) as our primary model. We also incorporate one-class learning with channel-robust training strategies to further learn a channel-invariant speech representation. Our submission achieved EER 20.33% in the DF task; EER 5.46% and min-tDCF 0.3094 in the LA task.

You Zhang, Ge Zhu, Fei Jiang et al.

Spoofing countermeasure (CM) systems are critical in speaker verification; they aim to discern spoofing attacks from bona fide speech trials. In practice, however, acoustic condition variability in speech utterances may significantly degrade the performance of CM systems. In this paper, we conduct a cross-dataset study on several state-of-the-art CM systems and observe significant performance degradation compared with their single-dataset performance. Observing differences of average magnitude spectra of bona fide utterances across the datasets, we hypothesize that channel mismatch among these datasets is one important reason. We then verify it by demonstrating a similar degradation of CM systems trained on original but evaluated on channel-shifted data. Finally, we propose several channel robust strategies (data augmentation, multi-task learning, adversarial learning) for CM systems, and observe a significant performance improvement on cross-dataset experiments.

Ge Zhu, Fei Jiang, Zhiyao Duan

State-of-the-art text-independent speaker verification systems typically use cepstral features or filter bank energies as speech features. Recent studies attempted to extract speaker embeddings directly from raw waveforms and have shown competitive results. In this paper, we propose a novel multi-scale waveform encoder that uses three convolution branches with different time scales to compute speech features from the waveform. These features are then processed by squeeze-and-excitation blocks, a multi-level feature aggregator, and a time delayed neural network (TDNN) to compute speaker embedding. We show that the proposed embeddings outperform existing raw-waveform-based speaker embeddings on speaker verification by a large margin. A further analysis of the learned filters shows that the multi-scale encoder attends to different frequency bands at its different scales while resulting in a more flat overall frequency response than any of the single-scale counterparts.