音频与语音处理：2025–2026 年度综述

An annual survey of the eess.AS arXiv category, covering research published from September 2025 – August 2026. Built on a close-reading set of 132 papers and organised into 37 chapters, it maps the year by the field's research sub-directions.

Book framework (share of the year's output):

• General Audio Event Detection and Separation (41.9%) — Focuses on robust sound classification, source separation, and detection in complex real-world acoustic scenes.
• Automatic Speech Recognition and Data Generation (25.4%) — Addresses ASR challenges including zero-shot adaptation, multilingual data synthesis, and low-latency conversational systems.
• Robust Text-to-Speech and Audio Enhancement (8.3%) — Improves generative speech synthesis and enhancement through noise robustness, emotional control, and adversarial defense.
• Real-time Spoken Dialogue and Deepfake Detection (8.0%) — Develops end-to-end spoken interaction pipelines, streaming recognition, and detection of synthetic or deepfake speech.
• Multimodal Audio-Visual Representation Learning (4.5%) — Explores audio-visual fusion, alignment, and encoding for robust recognition and generation tasks.
• Speech Quality Assessment and Evaluation Metrics (3.9%) — Establishes benchmarks and metrics for perceptual speech quality, MOS prediction, and automated speaking assessment.
• Discrete and Continuous Speech Token Modeling (3.7%) — Investigates neural codecs, discrete token representations, and continuous compression for efficient speech modeling.
• Self-Supervised Speech Representation Learning (1.9%) — Advances self-supervised learning techniques for extracting rich, robust representations from audio data.
• Music Generation and Audio Effects Synthesis (1.4%) — Covers AI-assisted music production, source restoration, and latent space manipulation for musical audio.
• Large Audio-Language Models and Reasoning (0.9%) — Examines reasoning capabilities, evaluation frameworks, and safety issues in large audio-language models.
• Specialized Audio Classification and Watermarking (0.1%) — Addresses niche tasks like few-shot incremental classification, multi-bit watermarking, and diffusion-based ASR decoding.

Each sub-direction is covered across three chapters — how it framed its questions, the strength of its key findings, and its open debates and frontiers — so the reader gets a traceable picture of what this part of the field actually produced over the year.