Yang and six co-authors survey brain-to-language decoding, tracing a field that moved from constrained recognition and acoustic reconstruction to text generation, streaming personalised speech and facial animation

The survey organises the task space of brain-to-language decoding into Articulated, Inner and Perceived tasks, and links each to the neural populations it engages, the representations available to decoders and the outputs those representations can support. Where earlier accounts were often organised by method or by signal modality, this synthesis uses task as the axis that ties neural populations, representations and outputs into one thread. Built on a search without a lower year limit and source-led updates through September 2026, covering invasive and non-invasive measurements.

The survey records that the field has expanded from constrained recognition and acoustic reconstruction to text generation, streaming personalised speech and facial animation. It treats the widening of output forms as a signature change in the field rather than staying on a single decoding-accuracy axis. Derived from synthesis of published work, with performance and communication costs compared within their reported protocols.

The survey identifies three complementary routes to progress: phonetic, acoustic and semantic targets preserve different aspects of a message; shared representations support reuse across recording conditions and tasks; and online communication increasingly depends on calibration, feedback and user control alongside decoding accuracy. It places evaluation and use-level factors such as calibration, feedback and user control alongside representation-level factors as conditions for progress, rather than treating accuracy as the sole metric. Synthesises algorithmic comparisons enabled by shared benchmarks together with longitudinal studies that reveal the demands of sustained use.

The survey outlines a prospective five-level trajectory from commands and language to meaning, scenarios and bidirectional cognitive exchange. It offers the field a layered long-term roadmap that discusses present capabilities and farther goals within one frame. This is a forward-looking outlook, presented alongside the preceding discussion of existing developments and their remaining limitations.

This survey is aimed at readers who want to understand brain-to-language decoding tasks, signals, methods, evaluation and practical use within a single frame; it covers invasive and non-invasive measurements, with a search without a lower year limit and source-led updates through September 2026. It connects Articulated, Inner and Perceived tasks to neural populations, available representations and supportable outputs, and compares published performance and communication costs within their reported protocols, so it is suited to locating where a given decoding target sits along which route, which representations it depends on, and what evaluation and use conditions it requires. The discussion of shared benchmarks and longitudinal studies offers reference points for reuse across recording conditions and tasks and for the demands of sustained use.

The loaded text is the arXiv abstract page and metadata, without body, figures or reference list, so specific performance values, the quantitative basis for communication costs, the individual entries of shared benchmarks and public resources, and the criteria for each of the five levels cannot be checked here. The search has no lower year limit and updates through September 2026, but its coverage and inclusion criteria are not expanded in the abstract. The complementary routes and the prospective trajectory are synthesis and outlook, and how testable they prove depends on later validation in shared benchmarks and longitudinal studies.