FunAudio-ASR Technical Report

In recent years, automatic speech recognition (ASR) has witnessed transformative advancements driven by three complementary paradigms: data scaling, model size scaling, and deep integration with large language models (LLMs). However, LLMs are prone to hallucination, which can significantly degrade user experience in real-world ASR applications. In this paper, we present FunAudio-ASR, a large-scale, LLM-based ASR system that synergistically combines massive data, large model capacity, LLM integration, and reinforcement learning to achieve state-of-the-art performance across diverse and complex speech recognition scenarios. Moreover, FunAudio-ASR is specifically optimized for practical deployment, with enhancements in streaming capability, noise robustness, code-switching, hotword customization, and satisfying other real-world application requirements. Experimental results show that while most LLM-based ASR systems achieve strong performance on open-source benchmarks, they often underperform on real industry evaluation sets. Thanks to production-oriented optimizations, FunAudio-ASR achieves SOTA performance on real application datasets, demonstrating its effectiveness and robustness in practical settings.

Insight Rumors: A Novel Textual Rumor Locating and Marking Model Leveraging Att_BiMamba2 Network

With the development of social media networks, rumor detection models have attracted more and more attention. Whereas, these models primarily focus on classifying contexts as rumors or not, lacking the capability to locate and mark specific rumor content. To address this limitation, this paper proposes a novel rumor detection model named Insight Rumors to locate and mark rumor content within textual data. Specifically, we propose the Bidirectional Mamba2 Network with Dot-Product Attention (Att_BiMamba2), a network that constructs a bidirectional Mamba2 model and applies dot-product attention to weight and combine the outputs from both directions, thereby enhancing the representation of high-dimensional rumor features. Simultaneously, a Rumor Locating and Marking module is designed to locate and mark rumors. The module constructs a skip-connection network to project high-dimensional rumor features onto low-dimensional label features. Moreover, Conditional Random Fields (CRF) is employed to impose strong constraints on the output label features, ensuring accurate rumor content location. Additionally, a labeled dataset for rumor locating and marking is constructed, with the effectiveness of the proposed model is evaluated through comprehensive experiments. Extensive experiments indicate that the proposed scheme not only detects rumors accurately but also locates and marks them in context precisely, outperforming state-of-the-art schemes that can only discriminate rumors roughly.

ClearerVoice-Studio: Bridging Advanced Speech Processing Research and Practical Deployment

This paper introduces ClearerVoice-Studio, an open-source, AI-powered speech processing toolkit designed to bridge cutting-edge research and practical application. Unlike broad platforms like SpeechBrain and ESPnet, ClearerVoice-Studio focuses on interconnected speech tasks of speech enhancement, separation, super-resolution, and multimodal target speaker extraction. A key advantage is its state-of-the-art pretrained models, including FRCRN with 3 million uses and MossFormer with 2.5 million uses, optimized for real-world scenarios. It also offers model optimization tools, multi-format audio support, the SpeechScore evaluation toolkit, and user-friendly interfaces, catering to researchers, developers, and end-users. Its rapid adoption attracting 3000 GitHub stars and 239 forks highlights its academic and industrial impact. This paper details ClearerVoice-Studio's capabilities, architectures, training strategies, benchmarks, community impact, and future plan. Source code is available at https://github.com/modelscope/ClearerVoice-Studio.

Plug-and-Play Co-Occurring Face Attention for Robust Audio-Visual Speaker Extraction

Audio-visual speaker extraction isolates a target speaker's speech from a mixture speech signal conditioned on a visual cue, typically using the target speaker's face recording. However, in real-world scenarios, other co-occurring faces are often present on-screen, providing valuable speaker activity cues in the scene. In this work, we introduce a plug-and-play inter-speaker attention module to process these flexible numbers of co-occurring faces, allowing for more accurate speaker extraction in complex multi-person environments. We integrate our module into two prominent models: the AV-DPRNN and the state-of-the-art AV-TFGridNet. Extensive experiments on diverse datasets, including the highly overlapped VoxCeleb2 and sparsely overlapped MISP, demonstrate that our approach consistently outperforms baselines. Furthermore, cross-dataset evaluations on LRS2 and LRS3 confirm the robustness and generalizability of our method.

ZenFlow: Enabling Stall-Free Offloading Training via Asynchronous Updates

Fine-tuning large language models (LLMs) often exceeds GPU memory limits, prompting systems to offload model states to CPU memory. However, existing offloaded training frameworks like ZeRO-Offload treat all parameters equally and update the full model on the CPU, causing severe GPU stalls, where fast, expensive GPUs sit idle waiting for slow CPU updates and limited-bandwidth PCIe transfers. We present ZenFlow, a new offloading framework that prioritizes important parameters and decouples updates between GPU and CPU. ZenFlow performs in-place updates of important gradients on GPU, while asynchronously offloading and accumulating less important ones on CPU, fully overlapping CPU work with GPU computation. To scale across GPUs, ZenFlow introduces a lightweight gradient selection method that exploits a novel spatial and temporal locality property of important gradients, avoiding costly global synchronization. ZenFlow achieves up to 5x end-to-end speedup, 2x lower PCIe traffic, and reduces GPU stalls by over 85 percent, all while preserving accuracy.

Multi-band Frequency Reconstruction for Neural Psychoacoustic Coding

Achieving high-fidelity audio compression while preserving perceptual quality across diverse content remains a key challenge in Neural Audio Coding (NAC). We introduce MUFFIN, a fully convolutional Neural Psychoacoustic Coding (NPC) framework that leverages psychoacoustically guided multi-band frequency reconstruction. At its core is a Multi-Band Spectral Residual Vector Quantization (MBS-RVQ) module that allocates bitrate across frequency bands based on perceptual salience. This design enables efficient compression while disentangling speaker identity from content using distinct codebooks. MUFFIN incorporates a transformer-inspired convolutional backbone and a modified snake activation to enhance resolution in fine-grained spectral regions. Experimental results on multiple benchmarks demonstrate that MUFFIN consistently outperforms existing approaches in reconstruction quality. A high-compression variant achieves a state-of-the-art 12.5 Hz rate with minimal loss. MUFFIN also proves effective in downstream generative tasks, highlighting its promise as a token representation for integration with language models. Audio samples and code are available.

Conditional Latent Diffusion-Based Speech Enhancement Via Dual Context Learning

Recently, the application of diffusion probabilistic models has advanced speech enhancement through generative approaches. However, existing diffusion-based methods have focused on the generation process in high-dimensional waveform or spectral domains, leading to increased generation complexity and slower inference speeds. Additionally, these methods have primarily modelled clean speech distributions, with limited exploration of noise distributions, thereby constraining the discriminative capability of diffusion models for speech enhancement. To address these issues, we propose a novel approach that integrates a conditional latent diffusion model (cLDM) with dual-context learning (DCL). Our method utilizes a variational autoencoder (VAE) to compress mel-spectrograms into a low-dimensional latent space. We then apply cLDM to transform the latent representations of both clean speech and background noise into Gaussian noise by the DCL process, and a parameterized model is trained to reverse this process, conditioned on noisy latent representations and text embeddings. By operating in a lower-dimensional space, the latent representations reduce the complexity of the generation process, while the DCL process enhances the model's ability to handle diverse and unseen noise environments. Our experiments demonstrate the strong performance of the proposed approach compared to existing diffusion-based methods, even with fewer iterative steps, and highlight the superior generalization capability of our models to out-of-domain noise datasets (https://github.com/modelscope/ClearerVoice-Studio).

HiFi-SR: A Unified Generative Transformer-Convolutional Adversarial Network for High-Fidelity Speech Super-Resolution

The application of generative adversarial networks (GANs) has recently advanced speech super-resolution (SR) based on intermediate representations like mel-spectrograms. However, existing SR methods that typically rely on independently trained and concatenated networks may lead to inconsistent representations and poor speech quality, especially in out-of-domain scenarios. In this work, we propose HiFi-SR, a unified network that leverages end-to-end adversarial training to achieve high-fidelity speech super-resolution. Our model features a unified transformer-convolutional generator designed to seamlessly handle both the prediction of latent representations and their conversion into time-domain waveforms. The transformer network serves as a powerful encoder, converting low-resolution mel-spectrograms into latent space representations, while the convolutional network upscales these representations into high-resolution waveforms. To enhance high-frequency fidelity, we incorporate a multi-band, multi-scale time-frequency discriminator, along with a multi-scale mel-reconstruction loss in the adversarial training process. HiFi-SR is versatile, capable of upscaling any input speech signal between 4 kHz and 32 kHz to a 48 kHz sampling rate. Experimental results demonstrate that HiFi-SR significantly outperforms existing speech SR methods across both objective metrics and ABX preference tests, for both in-domain and out-of-domain scenarios (https://github.com/modelscope/ClearerVoice-Studio).

MinMo: A Multimodal Large Language Model for Seamless Voice Interaction

Recent advancements in large language models (LLMs) and multimodal speech-text models have laid the groundwork for seamless voice interactions, enabling real-time, natural, and human-like conversations. Previous models for voice interactions are categorized as native and aligned. Native models integrate speech and text processing in one framework but struggle with issues like differing sequence lengths and insufficient pre-training. Aligned models maintain text LLM capabilities but are often limited by small datasets and a narrow focus on speech tasks. In this work, we introduce MinMo, a Multimodal Large Language Model with approximately 8B parameters for seamless voice interaction. We address the main limitations of prior aligned multimodal models. We train MinMo through multiple stages of speech-to-text alignment, text-to-speech alignment, speech-to-speech alignment, and duplex interaction alignment, on 1.4 million hours of diverse speech data and a broad range of speech tasks. After the multi-stage training, MinMo achieves state-of-the-art performance across various benchmarks for voice comprehension and generation while maintaining the capabilities of text LLMs, and also facilitates full-duplex conversation, that is, simultaneous two-way communication between the user and the system. Moreover, we propose a novel and simple voice decoder that outperforms prior models in voice generation. The enhanced instruction-following capabilities of MinMo supports controlling speech generation based on user instructions, with various nuances including emotions, dialects, and speaking rates, and mimicking specific voices. For MinMo, the speech-to-text latency is approximately 100ms, full-duplex latency is approximately 600ms in theory and 800ms in practice. The MinMo project web page is https://funaudiollm.github.io/minmo, and the code and models will be released soon.

Speech Separation using Neural Audio Codecs with Embedding Loss

Neural audio codecs have revolutionized audio processing by enabling speech tasks to be performed on highly compressed representations. Recent work has shown that speech separation can be achieved within these compressed domains, offering faster training and reduced inference costs. However, current approaches still rely on waveform-based loss functions, necessitating unnecessary decoding steps during training. We propose a novel embedding loss for neural audio codec-based speech separation that operates directly on compressed audio representations, eliminating the need for decoding during training. To validate our approach, we conduct comprehensive evaluations using both objective metrics and perceptual assessment techniques, including intrusive and non-intrusive methods. Our results demonstrate that embedding loss can be used to train codec-based speech separation models with a 2x improvement in training speed and computational cost while achieving better DNSMOS and STOI performance on the WSJ0-2mix dataset across 3 different pre-trained codecs.