声音与音频计算：2025–2026 年度综述

An annual survey of the cs.SD 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):

• Self-Supervised Audio Representation Learning (17.5%) — Focuses on improving general-purpose audio representations through self-supervised learning techniques like contrastive learning and contextual prediction for downstream classification tasks.
• Multimodal Audio-Video Generation and Synchronization (17.2%) — Explores joint generation and synchronization of audio with visual content, including lip-syncing, Foley generation, and spatial audio from video inputs.
• Speech Enhancement and Synthesis Systems (16.4%) — Develops neural models for high-fidelity speech synthesis, enhancement, and watermarking, often integrating text or visual priors for improved quality.
• Symbolic Music Generation and Analysis (10.4%) — Covers diffusion-based music generation, source separation, and dataset creation for symbolic music, addressing issues like de-duplication and evaluation metrics.
• Automatic Speech Recognition and Emotion Recognition (7.8%) — Enhances ASR robustness and emotion recognition capabilities using parameter-efficient fine-tuning, pseudo-labeling, and interpretability techniques.
• Zero-Shot Voice Conversion and Singing Synthesis (7.8%) — Investigates zero-shot voice identity morphing, singing voice separation, and timbre control using disentangled embeddings and latent diffusion models.
• Audio Deepfake Detection and Spoofing Defense (6.5%) — Addresses the detection of synthetic speech and audio spoofing attacks, focusing on generalizability, explainability, and robustness against evolving deepfake techniques.
• Large Audio-Language Models and Reasoning (6.3%) — Explores the training, fine-tuning, and safety guardrails of large language models adapted for audio understanding, reasoning, and multimodal interaction.
• Real-Time Audio Processing and Low-Latency Systems (4.8%) — Optimizes algorithms for real-time applications, including low-latency speech separation, streaming TTS, and efficient dynamic time warping implementations.
• End-to-End Dialogue and Speaker Diarization (3.4%) — Focuses on end-to-end frameworks for multi-speaker dialogue, speaker diarization, and spoken language understanding in complex acoustic environments.
• ASR and TTS Technical Reports and Benchmarks (1.9%) — Provides technical reports, benchmarks, and evaluation methodologies for automatic speech recognition and text-to-speech systems, including hallucination analysis.

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.