Yue Gu

Keyu An, Qian Chen, Chong Deng et al.

This report introduces FunAudioLLM, a model family designed to enhance natural voice interactions between humans and large language models (LLMs). At its core are two innovative models: SenseVoice, which handles multilingual speech recognition, emotion recognition, and audio event detection; and CosyVoice, which facilitates natural speech generation with control over multiple languages, timbre, speaking style, and speaker identity. SenseVoice-Small delivers exceptionally low-latency ASR for 5 languages, and SenseVoice-Large supports high-precision ASR for over 50 languages, while CosyVoice excels in multi-lingual voice generation, zero-shot in-context learning, cross-lingual voice cloning, and instruction-following capabilities. The models related to SenseVoice and CosyVoice have been open-sourced on Modelscope and Huggingface, along with the corresponding training, inference, and fine-tuning codes released on GitHub. By integrating these models with LLMs, FunAudioLLM enables applications such as speech-to-speech translation, emotional voice chat, interactive podcasts, and expressive audiobook narration, thereby pushing the boundaries of voice interaction technology. Demos are available at https://fun-audio-llm.github.io, and the code can be accessed at https://github.com/FunAudioLLM.

Amazon AGI, Aaron Langford, Aayush Shah et al. · amazon-science

We present Amazon Nova, a new generation of state-of-the-art foundation models that deliver frontier intelligence and industry-leading price performance. Amazon Nova Pro is a highly-capable multimodal model with the best combination of accuracy, speed, and cost for a wide range of tasks. Amazon Nova Lite is a low-cost multimodal model that is lightning fast for processing images, video, documents and text. Amazon Nova Micro is a text-only model that delivers our lowest-latency responses at very low cost. Amazon Nova Canvas is an image generation model that creates professional grade images with rich customization controls. Amazon Nova Reel is a video generation model offering high-quality outputs, customization, and motion control. Our models were built responsibly and with a commitment to customer trust, security, and reliability. We report benchmarking results for core capabilities, agentic performance, long context, functional adaptation, runtime performance, and human evaluation.

Zhihao Du, Qian Chen, Shiliang Zhang et al.

Recent years have witnessed a trend that large language model (LLM) based text-to-speech (TTS) emerges into the mainstream due to their high naturalness and zero-shot capacity. In this paradigm, speech signals are discretized into token sequences, which are modeled by an LLM with text as prompts and reconstructed by a token-based vocoder to waveforms. Obviously, speech tokens play a critical role in LLM-based TTS models. Current speech tokens are learned in an unsupervised manner, which lacks explicit semantic information and alignment to the text. In this paper, we propose to represent speech with supervised semantic tokens, which are derived from a multilingual speech recognition model by inserting vector quantization into the encoder. Based on the tokens, we further propose a scalable zero-shot TTS synthesizer, CosyVoice, which consists of an LLM for text-to-token generation and a conditional flow matching model for token-to-speech synthesis. Experimental results show that supervised semantic tokens significantly outperform existing unsupervised tokens in terms of content consistency and speaker similarity for zero-shot voice cloning. Moreover, we find that utilizing large-scale data further improves the synthesis performance, indicating the scalable capacity of CosyVoice. To the best of our knowledge, this is the first attempt to involve supervised speech tokens into TTS models.

Chenyang Gao, Yue Gu, Francesco Caliva et al.

Self-supervised speech representation learning (S3RL) is revolutionizing the way we leverage the ever-growing availability of data. While S3RL related studies typically use large models, we employ light-weight networks to comply with tight memory of compute-constrained devices. We demonstrate the effectiveness of S3RL on a keyword-spotting (KS) problem by using transformers with 330k parameters and propose a mechanism to enhance utterance-wise distinction, which proves crucial for improving performance on classification tasks. On the Google speech commands v2 dataset, the proposed method applied to the Auto-Regressive Predictive Coding S3RL led to a 1.2% accuracy improvement compared to training from scratch. On an in-house KS dataset with four different keywords, it provided 6% to 23.7% relative false accept improvement at fixed false reject rate. We argue this demonstrates the applicability of S3RL approaches to light-weight models for KS and confirms S3RL is a powerful alternative to traditional supervised learning for resource-constrained applications.

Gene-Ping Yang, Yue Gu, Qingming Tang et al.

Large self-supervised models are effective feature extractors, but their application is challenging under on-device budget constraints and biased dataset collection, especially in keyword spotting. To address this, we proposed a knowledge distillation-based self-supervised speech representation learning (S3RL) architecture for on-device keyword spotting. Our approach used a teacher-student framework to transfer knowledge from a larger, more complex model to a smaller, light-weight model using dual-view cross-correlation distillation and the teacher's codebook as learning objectives. We evaluated our model's performance on an Alexa keyword spotting detection task using a 16.6k-hour in-house dataset. Our technique showed exceptional performance in normal and noisy conditions, demonstrating the efficacy of knowledge distillation methods in constructing self-supervised models for keyword spotting tasks while working within on-device resource constraints.

Chenyang Gao, Yue Gu, Ivan Marsic

In supervised speech separation, permutation invariant training (PIT) is widely used to handle label ambiguity by selecting the best permutation to update the model. Despite its success, previous studies showed that PIT is plagued by excessive label assignment switching in adjacent epochs, impeding the model to learn better label assignments. To address this issue, we propose a novel training strategy, dynamic sample dropout (DSD), which considers previous best label assignments and evaluation metrics to exclude the samples that may negatively impact the learned label assignments during training. Additionally, we include layer-wise optimization (LO) to improve the performance by solving layer-decoupling. Our experiments showed that combining DSD and LO outperforms the baseline and solves excessive label assignment switching and layer-decoupling issues. The proposed DSD and LO approach is easy to implement, requires no extra training sets or steps, and shows generality to various speech separation tasks.

Zhihao Du, Yuxuan Wang, Qian Chen et al.

In our previous work, we introduced CosyVoice, a multilingual speech synthesis model based on supervised discrete speech tokens. By employing progressive semantic decoding with two popular generative models, language models (LMs) and Flow Matching, CosyVoice demonstrated high prosody naturalness, content consistency, and speaker similarity in speech in-context learning. Recently, significant progress has been made in multi-modal large language models (LLMs), where the response latency and real-time factor of speech synthesis play a crucial role in the interactive experience. Therefore, in this report, we present an improved streaming speech synthesis model, CosyVoice 2, which incorporates comprehensive and systematic optimizations. Specifically, we introduce finite-scalar quantization to improve the codebook utilization of speech tokens. For the text-speech LM, we streamline the model architecture to allow direct use of a pre-trained LLM as the backbone. In addition, we develop a chunk-aware causal flow matching model to support various synthesis scenarios, enabling both streaming and non-streaming synthesis within a single model. By training on a large-scale multilingual dataset, CosyVoice 2 achieves human-parity naturalness, minimal response latency, and virtually lossless synthesis quality in the streaming mode. We invite readers to listen to the demos at https://funaudiollm.github.io/cosyvoice2.

Jiayi Zhang, Chenxin Sun, Yue Gu et al.

The gaming industry has experienced substantial growth, but cheating in online games poses a significant threat to the integrity of the gaming experience. Cheating, particularly in first-person shooter (FPS) games, can lead to substantial losses for the game industry. Existing anti-cheat solutions have limitations, such as client-side hardware constraints, security risks, server-side unreliable methods, and both-sides suffer from a lack of comprehensive real-world datasets. To address these limitations, the paper proposes HAWK, a server-side FPS anti-cheat framework for the popular game CS:GO. HAWK utilizes machine learning techniques to mimic human experts' identification process, leverages novel multi-view features, and it is equipped with a well-defined workflow. The authors evaluate HAWK with the first large and real-world datasets containing multiple cheat types and cheating sophistication, and it exhibits promising efficiency and acceptable overheads, shorter ban times compared to the in-use anti-cheat, a significant reduction in manual labor, and the ability to capture cheaters who evaded official inspections.

Zhihao Du, Changfeng Gao, Yuxuan Wang et al.

In our prior works, we introduced a scalable streaming speech synthesis model, CosyVoice 2, which integrates a large language model (LLM) and a chunk-aware flow matching (FM) model, and achieves low-latency bi-streaming speech synthesis and human-parity quality. Despite these advancements, CosyVoice 2 exhibits limitations in language coverage, domain diversity, data volume, text formats, and post-training techniques. In this paper, we present CosyVoice 3, an improved model designed for zero-shot multilingual speech synthesis in the wild, surpassing its predecessor in content consistency, speaker similarity, and prosody naturalness. Key features of CosyVoice 3 include: 1) A novel speech tokenizer to improve prosody naturalness, developed via supervised multi-task training, including automatic speech recognition, speech emotion recognition, language identification, audio event detection, and speaker analysis. 2) A new differentiable reward model for post-training applicable not only to CosyVoice 3 but also to other LLM-based speech synthesis models. 3) Dataset Size Scaling: Training data is expanded from ten thousand hours to one million hours, encompassing 9 languages and 18 Chinese dialects across various domains and text formats. 4) Model Size Scaling: Model parameters are increased from 0.5 billion to 1.5 billion, resulting in enhanced performance on our multilingual benchmark due to the larger model capacity. These advancements contribute significantly to the progress of speech synthesis in the wild. We encourage readers to listen to the demo at https://funaudiollm.github.io/cosyvoice3.

Yue Gu, Zhihao Du, Ying Shi et al.

Recently, cross-attention-based contextual automatic speech recognition (ASR) models have made notable advancements in recognizing personalized biasing phrases. However, the effectiveness of cross-attention is affected by variations in biasing information volume, especially when the length of the biasing list increases significantly. We find that, regardless of the length of the biasing list, only a limited amount of biasing information is most relevant to a specific ASR intermediate representation. Therefore, by identifying and integrating the most relevant biasing information rather than the entire biasing list, we can alleviate the effects of variations in biasing information volume for contextual ASR. To this end, we propose a purified semantic correlation joint modeling (PSC-Joint) approach. In PSC-Joint, we define and calculate three semantic correlations between the ASR intermediate representations and biasing information from coarse to fine: list-level, phrase-level, and token-level. Then, the three correlations are jointly modeled to produce their intersection, so that the most relevant biasing information across various granularities is highlighted and integrated for contextual recognition. In addition, to reduce the computational cost introduced by the joint modeling of three semantic correlations, we also propose a purification mechanism based on a grouped-and-competitive strategy to filter out irrelevant biasing phrases. Compared with baselines, our PSC-Joint approach achieves average relative F1 score improvements of up to 21.34% on AISHELL-1 and 28.46% on KeSpeech, across biasing lists of varying lengths.

Chenyang Gao, Yue Gu, Ivan Marsic

Speech separation has been studied in time domain because of lower latency and higher performance compared to time-frequency domain. The masking-based method has been mostly used in time domain, and the other common method (mapping-based) has been inadequately studied. We investigate the use of the mapping-based method in the time domain and show that it can perform better on a large training set than the masking-based method. We also investigate the frequent label-switching problem in permutation invariant training (PIT), which results in suboptimal training because the labels selected by PIT differ across training epochs. Our experiment results showed that PIT works well in a shallow separation model, and the label switching occurs for a deeper model. We inferred that layer decoupling may be the reason for the frequent label switching. Therefore, we propose a training strategy based on progressive learning. This approach significantly reduced inconsistent label assignment without added computational complexity or training corpus. By combining this training strategy with the mapping-based method, we significantly improved the separation performance compared to the baseline.

Yue Gu, Wenxi Liu

In recent years, with the progress of deep learning technologies, crowd counting has been rapidly developed. In this work, we propose a simple yet effective crowd counting framework that is able to achieve the state-of-the-art performance on various crowded scenes. In particular, we first introduce a perspective-aware density map generation method that is able to produce ground-truth density maps from point annotations to train crowd counting model to accomplish superior performance than prior density map generation techniques. Besides, leveraging our density map generation method, we propose an iterative distillation algorithm to progressively enhance our model with identical network structures, without significantly sacrificing the dimension of the output density maps. In experiments, we demonstrate that, with our simple convolutional neural network architecture strengthened by our proposed training algorithm, our model is able to outperform or be comparable with the state-of-the-art methods. Furthermore, we also evaluate our density map generation approach and distillation algorithm in ablation studies.

Yue Gu, Zhihao Du, Hui Zhang et al.

Robustness against noise is critical for keyword spotting (KWS) in real-world environments. To improve the robustness, a speech enhancement front-end is involved. Instead of treating the speech enhancement as a separated preprocessing before the KWS system, in this study, a pre-trained speech enhancement front-end and a convolutional neural networks (CNNs) based KWS system are concatenated, where a feature transformation block is used to transform the output from the enhancement front-end into the KWS system's input. The whole model is trained jointly, thus the linguistic and other useful information from the KWS system can be back-propagated to the enhancement front-end to improve its performance. To fit the small-footprint device, a novel convolution recurrent network is proposed, which needs fewer parameters and computation and does not degrade performance. Furthermore, by changing the input features from the power spectrogram to Mel-spectrogram, less computation and better performance are obtained. our experimental results demonstrate that the proposed method significantly improves the KWS system with respect to noise robustness.

Jalal Abdulbaqi, Yue Gu, Ivan Marsic

Most current speech enhancement models use spectrogram features that require an expensive transformation and result in phase information loss. Previous work has overcome these issues by using convolutional networks to learn long-range temporal correlations across high-resolution waveforms. These models, however, are limited by memory-intensive dilated convolution and aliasing artifacts from upsampling. We introduce an end-to-end fully-recurrent hourglass-shaped neural network architecture with residual connections for waveform-based single-channel speech enhancement. Our model can efficiently capture long-range temporal dependencies by reducing the features resolution without information loss. Experimental results show that our model outperforms state-of-the-art approaches in six evaluation metrics.

Yue Gu, Kangning Yang, Shiyu Fu et al.

Multimodal affective computing, learning to recognize and interpret human affects and subjective information from multiple data sources, is still challenging because: (i) it is hard to extract informative features to represent human affects from heterogeneous inputs; (ii) current fusion strategies only fuse different modalities at abstract level, ignoring time-dependent interactions between modalities. Addressing such issues, we introduce a hierarchical multimodal architecture with attention and word-level fusion to classify utter-ance-level sentiment and emotion from text and audio data. Our introduced model outperforms the state-of-the-art approaches on published datasets and we demonstrated that our model is able to visualize and interpret the synchronized attention over modalities.

Yue Gu, Shuhong Chen, Ivan Marsic

In this paper, we present a novel deep multimodal framework to predict human emotions based on sentence-level spoken language. Our architecture has two distinctive characteristics. First, it extracts the high-level features from both text and audio via a hybrid deep multimodal structure, which considers the spatial information from text, temporal information from audio, and high-level associations from low-level handcrafted features. Second, we fuse all features by using a three-layer deep neural network to learn the correlations across modalities and train the feature extraction and fusion modules together, allowing optimal global fine-tuning of the entire structure. We evaluated the proposed framework on the IEMOCAP dataset. Our result shows promising performance, achieving 60.4% in weighted accuracy for five emotion categories.

Xinyu Li, Yanyi Zhang, Jianyu Zhang et al.

Process modeling and understanding are fundamental for advanced human-computer interfaces and automation systems. Most recent research has focused on activity recognition, but little has been done on sensor-based detection of process progress. We introduce a real-time, sensor-based system for modeling, recognizing and estimating the progress of a work process. We implemented a multimodal deep learning structure to extract the relevant spatio-temporal features from multiple sensory inputs and used a novel deep regression structure for overall completeness estimation. Using process completeness estimation with a Gaussian mixture model, our system can predict the phase for sequential processes. The performance speed, calculated using completeness estimation, allows online estimation of the remaining time. To train our system, we introduced a novel rectified hyperbolic tangent (rtanh) activation function and conditional loss. Our system was tested on data obtained from the medical process (trauma resuscitation) and sports events (Olympic swimming competition). Our system outperformed the existing trauma-resuscitation phase detectors with a phase detection accuracy of over 86%, an F1-score of 0.67, a completeness estimation error of under 12.6%, and a remaining-time estimation error of less than 7.5 minutes. For the Olympic swimming dataset, our system achieved an accuracy of 88%, an F1-score of 0.58, a completeness estimation error of 6.3% and a remaining-time estimation error of 2.9 minutes.