AlbEmo: A perceptually validated emotional speech dataset for Albanian

AlbEmo is, to our knowledge, the first emotional speech corpus for Albanian, a language covered by no current multilingual speech-emotion benchmark. Ten native Albanian speakers each produced six semantically neutral sentences in seven acted emotional styles (Ekman's six basic emotions, anger, disgust, fear, happiness, sadness, and surprise, plus a neutral reference), yielding 420 single-utterance recordings (48 kHz, 16-bit, mono WAV). Each speaker recorded alone through a purpose-built offline browser tool with a cardioid USB microphone, one utterance at a time. The emotional content was validated with a forced-choice perceptual listening test in which 1260 native-listener judgements give an overall recognition rate of 73% against a chance level of 14%, and Fleiss' $\kappa =0.59$ , in the range reported for comparable acted corpora. As a reference point for machine use, a baseline built on frozen self-supervised speech embeddings (WavLM) separates the seven emotions at 74% accuracy per speaker. The corpus and the anonymized listening-test responses are released openly under CC-BY-4.0, extending speech-emotion resources to an under-resourced Balkan language.