MeanVC, a streaming zero-shot voice conversion approach, uses a diffusion transformer with chunk-wise AR denoising and mean flows for efficient processing, achieving higher conversion quality with fewer parameters.

DialoSpeech combines an LLM with Chunked Flow Matching for dual-speaker dialogue speech synthesis, generating natural multi-turn conversations with speaker turns and overlaps in both Chinese and English.

WEALY uses Whisper decoder embeddings for audio-based lyrics matching, achieving comparable performance to state-of-the-art methods, while providing a reproducible pipeline.

Insertion Hallucination (IH), a phenomenon of generating non-existent sound events in video-to-audio, is addressed with a systematic evaluation framework and the Posterior Feature Correction (PFC) method, reducing IH by over 50%.

A framework for generative music infrastructure introduces inference-time attribution for verifiable artist compensation when their work conditions a generated output.

Pseudo2Real corrects pseudo-labels in ASR by fine-tuning two models, one on ground truth and the other on pseudo-labels, then applying their weight difference as a correction vector.

A multi-feature fusion approach, incorporating co-attention and personality characteristics, enhances multimodal depression detection in elderly, using LLDs, MFCCs, Wav2Vec, OpenFace, ResNet, and DenseNet features.

IntMeanFlow, with integral velocity distillation, achieves few-step speech generation by approximating average velocity with instantaneous velocity, eliminating Jacobian-vector products; O3S algorithm improves speech synthesis without overhead.

Bloodroot introduces a Watermark-as-Trigger concept in a backdoor framework using adversarial LoRA fine-tuning for stealthy audio poisoning; effective on speech recognition and speaker identification, remaining effective even under acoustic filtering.

Guitar tone morphing is achieved through LoRA fine-tuning and spherical interpolation using Music2Latent; the latter yields better tone transitions.

ACMID automatically creates a 7-stem music source separation dataset via web crawling and instrument classification using a pre-trained audio encoder; cleaning improves SDR performance by 2.39dB.

Full-Duplex-Bench-v2, a streaming framework with an automated examiner, evaluates full-duplex systems across daily, correction, entity tracking, and safety tasks; reveals confusion in simultaneous speech and struggles in correction handling.

SALAD-VAE, a semantic audio variational autoencoder, compresses audio at a low latent frame rate (7.8 Hz) using contrastive learning and CLAP-based embedding distillation; achieves SOTA compression and matches reconstruction quality of other VAEs.

INFER uses a frequency-domain neural framework conditioned on source/receiver positions to learn frequency response fields in confined spaces; outperforms time- and hybrid-domain baselines in car cabins.

A neural sound-matching approach leverages modulation extraction, constrained control signal parameterizations, and differentiable digital signal processing (DDSP) to discover modulations present in a sound.

TokenChain, a fully discrete speech chain coupling semantic-token ASR with a two-stage TTS, enables end-to-end feedback via straight-through argmax/Gumbel-Softmax, surpassing baseline accuracy and yielding better MOS.

EmoHRNet adapts High-Resolution Networks (HRNet) for speech emotion recognition by maintaining high-resolution spectrogram representations, achieving accuracies of 92.45% on benchmark datasets.

ECTSpeech incorporates Easy Consistency Tuning into speech synthesis, tightening consistency constraints on a pre-trained diffusion model and using a multi-scale gate module to achieve high-quality one-step generation.

Speech emotion recognition models trained on individual annotator ratings and optimized with standard SER systems, retaining all emotional ratings with soft-label distributions and incorporating penalization matrix show improved performance.

Statistical testing demonstrates that MFCCs inherently carry valuable prosodic information (energy, F0, voicing), challenging the assumption that they lack relevant temporal information.

The Segmented Full-Song Model (SFS) generates symbolic full-songs by factorizing a song into segments and generating each one through selective attention to related segments, achieving higher quality and efficiency.

LARA-Gen enables continuous emotion control in text-to-music models by aligning hidden states with an external music understanding model, achieving fine-grained control and improved emotion adherence and music quality.

FoleyGRAM aligns embeddings across video, text, and audio using the Gramian Representation Alignment Measure (GRAM) to semantically condition a diffusion-based audio synthesis model, enhancing video-to-audio generation.

StereoSync generates spatially-aware stereo audio from video by extracting spatial cues from depth maps and bounding boxes for cross-attention conditioning in a diffusion-based model, achieving temporal and spatial alignment.

EMORL-TTS, a reinforcement learning framework for LLM-based TTS, unifies global intensity control with local emphasis regulation, improving emotion accuracy, intensity differentiation, and emphasis clarity.

CxNet, a two-DNN system, separates monaural multi-speaker-image speech in reverberant conditions using a neural forward filtering module to estimate linear filters based on direct-path signals, improving reverberation capture.

A three-step framework transcribes guitar rhythmic patterns in polyphonic music using stem separation, MERT-based strum detection, and pattern decoding, achieving high accuracy and human-readable representation.

MSF-SER augments acoustic features with multi-granularity textual semantics (LES, GS, ES) fused via gated mechanisms and FiLM-modulated MOE, improving dimensional prediction on MSP-Podcast and IEMOCAP datasets.

Experiments on FACodec and Diff-HierVC discover a subspace within speaker embeddings whose removal alters machine perception but preserves human perception, enabling asynchronous voice anonymization with 100% success.

Sparse deepfake detection using TopK activation on AASIST embeddings improves detection performance (EER of 23.36%) and interpretability through latent representation analysis.

A reinforcement learning framework controls vocal tract articulator movements to generate syllable-level speech, achieving high similarity scores and accurate human transcription for syllables like "please" and "cat".

Sci-Phi, a spatial audio LLM, estimates parameters for sound sources and the environment using dual spatial and spectral encoders trained on synthetic Ambisonics recordings, demonstrating generalization to real room impulse responses.

AQA-TTRL, a test-time audio understanding framework, uses reinforcement learning with self-generated labels and confidence-based weighting to adapt LALMs on-the-fly, improving performance on MMAU, MMAR, and MMSU benchmarks.

WaveSP-Net combines prompt-tuning with wavelet transforms and a Mamba-based back-end for speech deepfake detection, outperforming state-of-the-art models on Deepfake-Eval-2024 and SpoofCeleb with low trainable parameters.

AUREXA-SE, an audio-visual speech enhancement framework, utilizes bidirectional cross-attention and Squeezeformer blocks for deep contextual fusion, achieving STOI of 0.516, PESQ of 1.323, and SI-SDR of -4.322 dB.

A non-probabilistic autoencoder with an independence constraint was developed for speech attribute conversion, yielding consistent signal transformation and demonstrating effectiveness on speaker identity and emotion.

A reproducible methodology evaluates preprocessing pipelines for in-the-wild TTS datasets using objective measures, acoustic descriptors, and speech-preservation metrics; identifies trade-offs between size, quality, and preservation.

CVSM uses contrastive self-supervised learning to learn music signal representations; maximizes similarity between vocal excerpts and musical mixtures; outperforms baselines in vocal similarity modeling tasks.

AudioToolAgent is a framework that uses a LLM agent to coordinate audio-language models as tools for question answering and speech-to-text; achieves state-of-the-art accuracy on multimodal datasets.

SALSA-V generates synchronized, high-fidelity long-form audio from silent video; masked diffusion and a shortcut loss enable fast, high-quality generation; outperforms existing methods in audio-visual alignment.

WavInWav uses a flow-based invertible neural network for time-domain speech hiding; a time-frequency loss on the time-domain signal improves secret audio quality during recovery; adds encryption for data protection.

Forensic Similarity determines if audio segments contain the same forensic traces using a two-part deep learning system; generalizes to unknown traces and is applicable to source verification and splicing detection.

Flamed-TTS, a zero-shot TTS framework, uses flow matching and discrete/continuous representations for efficient, high-fidelity speech synthesis; demonstrates improvements in intelligibility, naturalness, and speaker similarity.

Graph Neural Networks upscale low-resolution photogrammetry meshes for HRTF synthesis; validated geometrically and through perceptual evaluations; performance is compared to high-resolution scans and measured HRTFs.

SongFormer, a music structure analysis framework, learns from heterogeneous supervision; fuses audio representations and uses a source embedding; SongFormDB, a new corpus, supports scaling/evaluation.

STSM-FILM incorporates Feature-Wise Linear Modulation to condition a neural architecture for time-scale modification of speech; produces perceptually consistent outputs across time-scaling factors.

A four-stage pipeline produces detailed guitar tablature from audio; adapts piano transcription, classifies techniques, assigns strings/frets via transformer, and generates tablature with LSTM.

New methods for multi-source position and DOA estimation leverage Euclidean distance matrices; reduce computational cost compared to SRP-based methods; outperform SRP in two-source estimation accuracy.

A multi-view learning framework integrates contrastive principles into a generative pipeline; encodes audio latents into view-specific and view-common subspaces; improves performance on urban sound sensor network tasks.

HarmonicRNN, a linear RNN model, generates long music samples autoregressively; context-parallelism enables training on sequences up to one minute; achieves state-of-the-art results on small datasets.