A new paradigm called partially edited audio allows generative models to edit only selected portions of a real recording while leaving the rest untouched. This technique differs from full audio generation, which creates an entirely synthetic signal. Partially edited audio can correct mispronunciations, change background sounds, or insert new speech, but it can also be used to spread misinformation.

Early methods used simple splicing or copy‑move, which leave visible boundary glitches and mismatched background signatures. Modern context‑aware infilling replaces the spliced portion by blending it seamlessly with the surrounding audio, using autoregressive codec language models or non‑autoregressive diffusion in spectrogram space. These models predict acoustic tokens for a masked span conditioned on both its left and right context, achieving higher consistency.

Detecting such edits is challenging because the tampered region is often brief. Models trained only on fully generated fake audio perform poorly on partially edited data. However, detectors trained on short embedded fake portions generalise well across both fully and partially edited scenarios. Yet most defence research and benchmark datasets focus on single‑speaker, clean studio recordings and temporal edits, leaving paralinguistic, extralinguistic, and other attribute edits largely unexplored. Real‑world recordings are noisy, reverberant, and may contain multiple speakers, making current benchmarks less realistic.

Future defence strategies propose proactive accountability: embedding watermarks during generation so that later edits can be traced and verified. Adaptive detection pipelines combine self‑supervised learning, large‑language‑model‑based detectors, and watermark checks to provide traceability.

Resources such as PartialSpoof, HAD, and the ADD challenge series now include segment‑level localisation, multimodal forgeries, and instruction‑driven partial fakes. These datasets extend coverage beyond simple speech edits, but they still miss edits affecting prosody, emotion, or background.