오디오 캡션

Audio captioning connects an audio signal with free-form descriptive text. A caption may identify sound sources, qualities or relationships, while automatic speech recognition targets the words spoken in the signal. A description such as birds chirping over running water therefore has a different purpose from a verbatim speech transcript. Clotho and AudioCaps provide published datasets for studying this task; naming them does not establish that they are the only benchmarks or always the most widely used. The original Clotho paper reports 4981 audio samples lasting 15 to 30 seconds and five caption references per file. Multiple references support varied descriptions of the same content rather than requiring one identical sentence from every annotator. Its baseline uses an encoder-decoder scheme with attention, supplied with 64 log-mel-band features. The encoder maps the feature sequence into representations that the decoder uses to produce words. This particular baseline does not establish the architecture of every later audio-language system. The paper evaluates generated captions with several reference-comparison metrics, including CIDEr. CIDEr compares weighted textual n-gram information and should not be confused with a direct measure of acoustic frequency, loudness or proof that every mentioned event is present. Grounded description also needs inspection of the actual recording. Non-speech captions can supply information that speech subtitles omit, offering a possible accessibility use to evaluate with intended readers. Fluent wording remains separate from correct identification of events, and an audio-only system should not invent unsupported visual attributes or scene details.