Zeyu Jin

Joseph Turian, Jordie Shier, Humair Raj Khan et al. · cmu

What audio embedding approach generalizes best to a wide range of downstream tasks across a variety of everyday domains without fine-tuning? The aim of the HEAR benchmark is to develop a general-purpose audio representation that provides a strong basis for learning in a wide variety of tasks and scenarios. HEAR evaluates audio representations using a benchmark suite across a variety of domains, including speech, environmental sound, and music. HEAR was launched as a NeurIPS 2021 shared challenge. In the spirit of shared exchange, each participant submitted an audio embedding model following a common API that is general-purpose, open-source, and freely available to use. Twenty-nine models by thirteen external teams were evaluated on nineteen diverse downstream tasks derived from sixteen datasets. Open evaluation code, submitted models and datasets are key contributions, enabling comprehensive and reproducible evaluation, as well as previously impossible longitudinal studies. It still remains an open question whether one single general-purpose audio representation can perform as holistically as the human ear.

Nikhil Kandpal, Oriol Nieto, Zeyu Jin

Consumer-grade music recordings such as those captured by mobile devices typically contain distortions in the form of background noise, reverb, and microphone-induced EQ. This paper presents a deep learning approach to enhance low-quality music recordings by combining (i) an image-to-image translation model for manipulating audio in its mel-spectrogram representation and (ii) a music vocoding model for mapping synthetically generated mel-spectrograms to perceptually realistic waveforms. We find that this approach to music enhancement outperforms baselines which use classical methods for mel-spectrogram inversion and an end-to-end approach directly mapping noisy waveforms to clean waveforms. Additionally, in evaluating the proposed method with a listening test, we analyze the reliability of common audio enhancement evaluation metrics when used in the music domain.

Oriol Nieto, Zeyu Jin, Franck Dernoncourt et al.

Spoken language recognition (SLR) is the task of automatically identifying the language present in a speech signal. Existing SLR models are either too computationally expensive or too large to run effectively on devices with limited resources. For real-world deployment, a model should also gracefully handle unseen languages outside of the target language set, yet prior work has focused on closed-set classification where all input languages are known a-priori. In this paper we address these two limitations: we explore efficient model architectures for SLR based on convolutional networks, and propose a multilabel training strategy to handle non-target languages at inference time. Using the VoxLingua107 dataset, we show that our models obtain competitive results while being orders of magnitude smaller and faster than current state-of-the-art methods, and that our multilabel strategy is more robust to unseen non-target languages compared to multiclass classification.

MiniMax, Aili Chen, Aonian Li et al.

We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.

Ke Chen, Jiaqi Su, Taylor Berg-Kirkpatrick et al.

Achieving robust speech separation for overlapping speakers in various acoustic environments with noise and reverberation remains an open challenge. Although existing datasets are available to train separators for specific scenarios, they do not effectively generalize across diverse real-world scenarios. In this paper, we present a novel data simulation pipeline that produces diverse training data from a range of acoustic environments and content, and propose new training paradigms to improve quality of a general speech separation model. Specifically, we first introduce AC-SIM, a data simulation pipeline that incorporates broad variations in both content and acoustics. Then we integrate multiple training objectives into the permutation invariant training (PIT) to enhance separation quality and generalization of the trained model. Finally, we conduct comprehensive objective and human listening experiments across separation architectures and benchmarks to validate our methods, demonstrating substantial improvement of generalization on both non-homologous and real-world test sets.

Sreyan Ghosh, Chandra Kiran Reddy Evuru, Sonal Kumar et al.

Instruction Tuning (IT), the process of training large language models (LLMs) using instruction-response pairs, has emerged as the predominant method for transforming base pre-trained LLMs into open-domain conversational agents. While IT has achieved notable success and widespread adoption, its limitations and shortcomings remain underexplored. In this paper, through rigorous experiments and an in-depth analysis of the changes LLMs undergo through IT, we reveal various limitations of IT. In particular, we show that (1) IT fails to enhance knowledge or skills in LLMs. LoRA fine-tuning is limited to learning response initiation and style tokens, and full-parameter fine-tuning leads to knowledge degradation. (2) Copying response patterns from IT datasets derived from knowledgeable sources leads to a decline in response quality. (3) Full-parameter fine-tuning increases hallucination by inaccurately borrowing tokens from conceptually similar instances in the IT dataset for generating responses. (4) Popular methods to improve IT do not lead to performance improvements over a simple LoRA fine-tuned model. Our findings reveal that responses generated solely from pre-trained knowledge consistently outperform responses by models that learn any form of new knowledge from IT on open-source datasets. We hope the insights and challenges revealed in this paper inspire future work in related directions.

Zeyu Jin, Jia Jia, Qixin Wang et al.

Speech-language multi-modal learning presents a significant challenge due to the fine nuanced information inherent in speech styles. Therefore, a large-scale dataset providing elaborate comprehension of speech style is urgently needed to facilitate insightful interplay between speech audio and natural language. However, constructing such datasets presents a major trade-off between large-scale data collection and high-quality annotation. To tackle this challenge, we propose an automatic speech annotation system for expressiveness interpretation that annotates in-the-wild speech clips with expressive and vivid human language descriptions. Initially, speech audios are processed by a series of expert classifiers and captioning models to capture diverse speech characteristics, followed by a fine-tuned LLaMA for customized annotation generation. Unlike previous tag/templet-based annotation frameworks with limited information and diversity, our system provides in-depth understandings of speech style through tailored natural language descriptions, thereby enabling accurate and voluminous data generation for large model training. With this system, we create SpeechCraft, a fine-grained bilingual expressive speech dataset. It is distinguished by highly descriptive natural language style prompts, containing approximately 2,000 hours of audio data and encompassing over two million speech clips. Extensive experiments demonstrate that the proposed dataset significantly boosts speech-language task performance in stylist speech synthesis and speech style understanding.

Zeyu Jin, Songtao Zhou, Haoyu Wang et al.

The recent advancement of Artificial Intelligence Generated Content (AIGC) has led to significant strides in modeling human interaction, particularly in the context of multimodal dialogue. While current methods impressively generate realistic dialogue in isolated modalities like speech or vision, challenges remain in controllable Multimodal Dialogue Generation (MDG). This paper focuses on the natural alignment between speech, vision, and text in human interaction, aiming for expressive dialogue generation through multimodal conditional control. To address the insufficient richness and diversity of dialogue expressiveness in existing datasets, we introduce a novel multimodal dialogue annotation pipeline to curate dialogues from movies and TV series with fine-grained annotations in interactional characteristics. The resulting MM-Dia dataset (360+ hours, 54,700 dialogues) facilitates explicitly controlled MDG, specifically through style-controllable dialogue speech synthesis. In parallel, MM-Dia-Bench (309 highly expressive dialogues with visible single-/dual-speaker scenes) serves as a rigorous testbed for implicit cross-modal MDG control, evaluating audio-visual style consistency across modalities. Extensive experiments demonstrate that training on MM-Dia significantly enhances fine-grained controllability, while evaluations on MM-Dia-Bench reveal limitations in current frameworks to replicate the nuanced expressiveness of human interaction. These findings provides new insights and challenges for multimodal conditional dialogue generation.

Sreyan Ghosh, Chandra Kiran Reddy Evuru, Sonal Kumar et al.

Large Vision-Language Models (LVLMs) often produce responses that misalign with factual information, a phenomenon known as hallucinations. While hallucinations are well-studied, the exact causes behind them remain underexplored. In this paper, we first investigate the root causes of hallucinations in LVLMs. Our findings reveal that existing mitigation techniques primarily reduce hallucinations for visual recognition prompts-those that require simple descriptions of visual elements-but fail for cognitive prompts that demand deliberate reasoning. We identify the core issue as a lack of true visual perception in LVLMs: although they can accurately recognize visual elements, they struggle to fully interpret these elements in the context of the input prompt and effectively link this recognition to their internal knowledge, which is critical for reasoning. To address this gap, we introduce Visual Description Grounded Decoding (VDGD), a simple, robust, and training-free method designed to enhance visual perception and improve reasoning capabilities in LVLMs. VDGD works by first generating a detailed description of the image and appending it as a prefix to the instruction. During response generation, tokens are sampled based on their KL divergence to the description, favoring candidates with lower divergence. Experimental results on multiple visual reasoning benchmarks and LVLMs demonstrate that VDGD consistently outperforms existing baselines 2% - 33%. Finally, we introduce VaLLu, a benchmark designed for comprehensive evaluation of the cognitive capabilities of LVLMs.

Yingahao Aaron Li, Rithesh Kumar, Zeyu Jin

Diffusion models have demonstrated significant potential in speech synthesis tasks, including text-to-speech (TTS) and voice cloning. However, their iterative denoising processes are computationally intensive, and previous distillation attempts have shown consistent quality degradation. Moreover, existing TTS approaches are limited by non-differentiable components or iterative sampling that prevent true end-to-end optimization with perceptual metrics. We introduce DMOSpeech, a distilled diffusion-based TTS model that uniquely achieves both faster inference and superior performance compared to its teacher model. By enabling direct gradient pathways to all model components, we demonstrate the first successful end-to-end optimization of differentiable metrics in TTS, incorporating Connectionist Temporal Classification (CTC) loss and Speaker Verification (SV) loss. Our comprehensive experiments, validated through extensive human evaluation, show significant improvements in naturalness, intelligibility, and speaker similarity while reducing inference time by orders of magnitude. This work establishes a new framework for aligning speech synthesis with human auditory preferences through direct metric optimization. The audio samples are available at https://dmospeech.github.io/.

Stephen Brade, Sam Anderson, Rithesh Kumar et al.

Novice content creators often invest significant time recording expressive speech for social media videos. While recent advancements in text-to-speech (TTS) technology can generate highly realistic speech in various languages and accents, many struggle with unintuitive or overly granular TTS interfaces. We propose simplifying TTS generation by allowing users to specify high-level context alongside their script. Our Wizard-of-Oz system, SpeakEasy, leverages user-provided context to inform and influence TTS output, enabling iterative refinement with high-level feedback. This approach was informed by two 8-subject formative studies: one examining content creators' experiences with TTS, and the other drawing on effective strategies from voice actors. Our evaluation shows that participants using SpeakEasy were more successful in generating performances matching their personal standards, without requiring significantly more effort than leading industry interfaces.

Zeyu Jin, Xiaoyu Qin, Songtao Zhou et al.

Soccer commentary plays a crucial role in enhancing the soccer game viewing experience for audiences. Previous studies in automatic soccer commentary generation typically adopt an end-to-end method to generate anonymous live text commentary. Such generated commentary is insufficient in the context of real-world live televised commentary, as it contains anonymous entities, context-dependent errors and lacks statistical insights of the game events. To bridge the gap, we propose GameSight, a two-stage model to address soccer commentary generation as a knowledge-enhanced visual reasoning task, enabling live-televised-like knowledgeable commentary with accurate reference to entities (players and teams). GameSight starts by performing visual reasoning to align anonymous entities with fine-grained visual and contextual analysis. Subsequently, the entity-aligned commentary is refined with knowledge by incorporating external historical statistics and iteratively updated internal game state information. Consequently, GameSight improves the player alignment accuracy by 18.5% on SN-Caption-test-align dataset compared to Gemini 2.5-pro. Combined with further knowledge enhancement, GameSight outperforms in segment-level accuracy and commentary quality, as well as game-level contextual relevance and structural composition. We believe that our work paves the way for a more informative and engaging human-centric experience with the AI sports application. Demo Page: https://gamesight2025.github.io/gamesight2025

Justin Lovelace, Rithesh Kumar, Jiaqi Su et al. · cmu

While generative Text-to-Speech (TTS) systems leverage vast ``in-the-wild" data to achieve remarkable success, speech-to-speech processing tasks like enhancement face data limitations, which lead data-hungry generative approaches to distort speech content and speaker identity. To bridge this gap, we present SpeechOp, a multi-task latent diffusion model that transforms pre-trained TTS models into a universal speech processor capable of performing a wide range of speech tasks and composing them in novel ways at inference time. By adapting a pre-trained TTS model, SpeechOp inherits a rich understanding of natural speech, accelerating training and improving S2S task quality, while simultaneously enhancing core TTS performance. Finally, we introduce Implicit Task Composition (ITC), a novel pipeline where ASR-derived transcripts (e.g., from Whisper) guide SpeechOp's enhancement via our principled inference-time task composition. ITC achieves state-of-the-art content preservation by robustly combining web-scale speech understanding with SpeechOp's generative capabilities. Audio samples are available at https://justinlovelace.github.io/projects/speechop

Max Morrison, Zeyu Jin, Nicholas J. Bryan et al.

Modifying the pitch and timing of an audio signal are fundamental audio editing operations with applications in speech manipulation, audio-visual synchronization, and singing voice editing and synthesis. Thus far, methods for pitch-shifting and time-stretching that use digital signal processing (DSP) have been favored over deep learning approaches due to their speed and relatively higher quality. However, even existing DSP-based methods for pitch-shifting and time-stretching induce artifacts that degrade audio quality. In this paper, we propose Controllable LPCNet (CLPCNet), an improved LPCNet vocoder capable of pitch-shifting and time-stretching of speech. For objective evaluation, we show that CLPCNet performs pitch-shifting of speech on unseen datasets with high accuracy relative to prior neural methods. For subjective evaluation, we demonstrate that the quality and naturalness of pitch-shifting and time-stretching with CLPCNet on unseen datasets meets or exceeds competitive neural- or DSP-based approaches.

Shuqi Dai, Zeyu Jin, Celso Gomes et al.

Recent advances in deep learning have expanded possibilities to generate music, but generating a customizable full piece of music with consistent long-term structure remains a challenge. This paper introduces MusicFrameworks, a hierarchical music structure representation and a multi-step generative process to create a full-length melody guided by long-term repetitive structure, chord, melodic contour, and rhythm constraints. We first organize the full melody with section and phrase-level structure. To generate melody in each phrase, we generate rhythm and basic melody using two separate transformer-based networks, and then generate the melody conditioned on the basic melody, rhythm and chords in an auto-regressive manner. By factoring music generation into sub-problems, our approach allows simpler models and requires less data. To customize or add variety, one can alter chords, basic melody, and rhythm structure in the music frameworks, letting our networks generate the melody accordingly. Additionally, we introduce new features to encode musical positional information, rhythm patterns, and melodic contours based on musical domain knowledge. A listening test reveals that melodies generated by our method are rated as good as or better than human-composed music in the POP909 dataset about half the time.

Max Morrison, Lucas Rencker, Zeyu Jin et al.

Text-based speech editors expedite the process of editing speech recordings by permitting editing via intuitive cut, copy, and paste operations on a speech transcript. A major drawback of current systems, however, is that edited recordings often sound unnatural because of prosody mismatches around edited regions. In our work, we propose a new context-aware method for more natural sounding text-based editing of speech. To do so, we 1) use a series of neural networks to generate salient prosody features that are dependent on the prosody of speech surrounding the edit and amenable to fine-grained user control 2) use the generated features to control a standard pitch-shift and time-stretch method and 3) apply a denoising neural network to remove artifacts induced by the signal manipulation to yield a high-fidelity result. We evaluate our approach using a subjective listening test, provide a detailed comparative analysis, and conclude several interesting insights.

Pranay Manocha, Zeyu Jin, Richard Zhang et al.

Many speech processing methods based on deep learning require an automatic and differentiable audio metric for the loss function. The DPAM approach of Manocha et al. learns a full-reference metric trained directly on human judgments, and thus correlates well with human perception. However, it requires a large number of human annotations and does not generalize well outside the range of perturbations on which it was trained. This paper introduces CDPAM, a metric that builds on and advances DPAM. The primary improvement is to combine contrastive learning and multi-dimensional representations to build robust models from limited data. In addition, we collect human judgments on triplet comparisons to improve generalization to a broader range of audio perturbations. CDPAM correlates well with human responses across nine varied datasets. We also show that adding this metric to existing speech synthesis and enhancement methods yields significant improvement, as measured by objective and subjective tests.

Jongpil Lee, Nicholas J. Bryan, Justin Salamon et al.

Deep representation learning offers a powerful paradigm for mapping input data onto an organized embedding space and is useful for many music information retrieval tasks. Two central methods for representation learning include deep metric learning and classification, both having the same goal of learning a representation that can generalize well across tasks. Along with generalization, the emerging concept of disentangled representations is also of great interest, where multiple semantic concepts (e.g., genre, mood, instrumentation) are learned jointly but remain separable in the learned representation space. In this paper we present a single representation learning framework that elucidates the relationship between metric learning, classification, and disentanglement in a holistic manner. For this, we (1) outline past work on the relationship between metric learning and classification, (2) extend this relationship to multi-label data by exploring three different learning approaches and their disentangled versions, and (3) evaluate all models on four tasks (training time, similarity retrieval, auto-tagging, and triplet prediction). We find that classification-based models are generally advantageous for training time, similarity retrieval, and auto-tagging, while deep metric learning exhibits better performance for triplet-prediction. Finally, we show that our proposed approach yields state-of-the-art results for music auto-tagging.

Jongpil Lee, Nicholas J. Bryan, Justin Salamon et al.

Music similarity search is useful for a variety of creative tasks such as replacing one music recording with another recording with a similar "feel", a common task in video editing. For this task, it is typically necessary to define a similarity metric to compare one recording to another. Music similarity, however, is hard to define and depends on multiple simultaneous notions of similarity (i.e. genre, mood, instrument, tempo). While prior work ignore this issue, we embrace this idea and introduce the concept of multidimensional similarity and unify both global and specialized similarity metrics into a single, semantically disentangled multidimensional similarity metric. To do so, we adapt a variant of deep metric learning called conditional similarity networks to the audio domain and extend it using track-based information to control the specificity of our model. We evaluate our method and show that our single, multidimensional model outperforms both specialized similarity spaces and alternative baselines. We also run a user-study and show that our approach is favored by human annotators as well.

Max Morrison, Zeyu Jin, Justin Salamon et al.

Speech synthesis has recently seen significant improvements in fidelity, driven by the advent of neural vocoders and neural prosody generators. However, these systems lack intuitive user controls over prosody, making them unable to rectify prosody errors (e.g., misplaced emphases and contextually inappropriate emotions) or generate prosodies with diverse speaker excitement levels and emotions. We address these limitations with a user-controllable, context-aware neural prosody generator. Given a real or synthesized speech recording, our model allows a user to input prosody constraints for certain time frames and generates the remaining time frames from input text and contextual prosody. We also propose a pitch-shifting neural vocoder to modify input speech to match the synthesized prosody. Through objective and subjective evaluations we show that we can successfully incorporate user control into our prosody generation model without sacrificing the overall naturalness of the synthesized speech.

Jiaqi Su, Zeyu Jin, Adam Finkelstein

Real-world audio recordings are often degraded by factors such as noise, reverberation, and equalization distortion. This paper introduces HiFi-GAN, a deep learning method to transform recorded speech to sound as though it had been recorded in a studio. We use an end-to-end feed-forward WaveNet architecture, trained with multi-scale adversarial discriminators in both the time domain and the time-frequency domain. It relies on the deep feature matching losses of the discriminators to improve the perceptual quality of enhanced speech. The proposed model generalizes well to new speakers, new speech content, and new environments. It significantly outperforms state-of-the-art baseline methods in both objective and subjective experiments.

Kaizhi Qian, Zeyu Jin, Mark Hasegawa-Johnson et al.

Non-parallel many-to-many voice conversion remains an interesting but challenging speech processing task. Many style-transfer-inspired methods such as generative adversarial networks (GANs) and variational autoencoders (VAEs) have been proposed. Recently, AutoVC, a conditional autoencoders (CAEs) based method achieved state-of-the-art results by disentangling the speaker identity and speech content using information-constraining bottlenecks, and it achieves zero-shot conversion by swapping in a different speaker's identity embedding to synthesize a new voice. However, we found that while speaker identity is disentangled from speech content, a significant amount of prosodic information, such as source F0, leaks through the bottleneck, causing target F0 to fluctuate unnaturally. Furthermore, AutoVC has no control of the converted F0 and thus unsuitable for many applications. In the paper, we modified and improved autoencoder-based voice conversion to disentangle content, F0, and speaker identity at the same time. Therefore, we can control the F0 contour, generate speech with F0 consistent with the target speaker, and significantly improve quality and similarity. We support our improvement through quantitative and qualitative analysis.

Pranay Manocha, Adam Finkelstein, Richard Zhang et al.

Many audio processing tasks require perceptual assessment. The ``gold standard`` of obtaining human judgments is time-consuming, expensive, and cannot be used as an optimization criterion. On the other hand, automated metrics are efficient to compute but often correlate poorly with human judgment, particularly for audio differences at the threshold of human detection. In this work, we construct a metric by fitting a deep neural network to a new large dataset of crowdsourced human judgments. Subjects are prompted to answer a straightforward, objective question: are two recordings identical or not? These pairs are algorithmically generated under a variety of perturbations, including noise, reverb, and compression artifacts; the perturbation space is probed with the goal of efficiently identifying the just-noticeable difference (JND) level of the subject. We show that the resulting learned metric is well-calibrated with human judgments, outperforming baseline methods. Since it is a deep network, the metric is differentiable, making it suitable as a loss function for other tasks. Thus, simply replacing an existing loss (e.g., deep feature loss) with our metric yields significant improvement in a denoising network, as measured by subjective pairwise comparison.

Ohad Fried, Ayush Tewari, Michael Zollhöfer et al.

Editing talking-head video to change the speech content or to remove filler words is challenging. We propose a novel method to edit talking-head video based on its transcript to produce a realistic output video in which the dialogue of the speaker has been modified, while maintaining a seamless audio-visual flow (i.e. no jump cuts). Our method automatically annotates an input talking-head video with phonemes, visemes, 3D face pose and geometry, reflectance, expression and scene illumination per frame. To edit a video, the user has to only edit the transcript, and an optimization strategy then chooses segments of the input corpus as base material. The annotated parameters corresponding to the selected segments are seamlessly stitched together and used to produce an intermediate video representation in which the lower half of the face is rendered with a parametric face model. Finally, a recurrent video generation network transforms this representation to a photorealistic video that matches the edited transcript. We demonstrate a large variety of edits, such as the addition, removal, and alteration of words, as well as convincing language translation and full sentence synthesis.