AlbEmo: a perceptually validated emotional speech corpus for Albanian

AlbEmo is, to our knowledge, the first emotional speech corpus for Albanian, a language not covered by current multilingual speech-emotion benchmarks. Ten native Albanian speakers, five female and five male, aged 18 to over 55, each produced six semantically neutral sentences in seven emotional styles: the six basic emotions of Ekman plus a neutral reference. The result is 420 recordings, one for every combination of speaker, emotion and sentence, with no missing cells. Audio is uncompressed WAV at 48 kHz, 16-bit, mono, totalling 22.1 minutes and about 127 MB. Clips average 3.15 seconds, ranging from 1.2 to 10.1 seconds. Perceptual validation. Every emotion label was validated in a blind forced-choice listening test. Four native-Albanian evaluators contributed 1,260 judgements, three independent ratings for each of the 420 clips, and never rated their own recordings. Overall recognition is 73% against a chance level of 14%, with a mean unbiased hit-rate (Wagner) of 54% and Fleiss' kappa of 0.59, which is in the range reported for comparable acted corpora. Every individual judgement is released, not only the aggregate, so users can weight or filter training examples by listener agreement. Recording. Speakers recorded alone in a quiet office at the University of Tirana with a Blue Yeti cardioid USB microphone, one utterance at a time, using a purpose-built offline browser tool included in the deposit. Browser-side processing (echo cancellation, noise suppression, automatic gain control) was disabled, so the audio is exactly as recorded, with no denoising or level normalization. No sample reaches full scale, so no file clips, and in the median recording the speech sits 28 dB above the background, never less than 17 dB in any file. Layout. The archive unpacks to a single albemo/ directory: recordings/spNN//sentence_K.wav — the audio, 10 speakers x 7 emotions x 6 sentences corpus-sentences.json — the six sentences and the seven emotion codes speaker-demographics.csv — gender, age band, region of origin, other languages manifest.csv — one row per recording: filename, speaker, emotion, sentence, duration, sample rate, bit depth, channels evaluations/ — 30 per-rater response files, aggregated results, and the integer confusion matrix scripts/ — the analysis scripts that reproduce the reported numbers tools/ — the offline browser recorder and listening-test tools, in Albanian and English consent-template.md, README.md, LICENSE, checksums.md5 spNN is the anonymized speaker, NN from 01 to 10; is the Albanian emotion marker; K from 1 to 6 indexes the sentence. For example, recordings/sp02/trishtim/sentence_4.wav is speaker 2 producing sentence S04 in a sad voice. Emotion markers, in the order recorded, with the internal code: neutral (neutral, e07), gëzim (happiness, e04), trishtim (sadness, e05), zemërim (anger, e01), frikë (fear, e03), habi (surprise, e06), neveri (disgust, e02). Loading. Each recording's emotion label is its parent directory, so the corpus loads without a separate label file: import soundfile as sf from pathlib import Path for f in Path("recordings").glob("sp*/*/sentence_*.wav"): emotion = f.parent.name # Albanian marker, e.g. trishtim audio, sr = sf.read(f) # sr = 48000 The per-response files in evaluations/ join to the audio on (speaker, emotion, sentence). Intended uses. Speech emotion recognition, including multilingual and cross-lingual work needing a language absent from current pretraining benchmarks; phonetic and prosodic study of Albanian; and transfer-learning and low-resource research. The included browser tools need no server and no installation, so the same collection protocol can be applied to another under-resourced language. Limitations. The emotions are portrayed rather than spontaneous, as in EmoDB, RAVDESS and SUBESCO, so results need not transfer directly to naturally occurring emotional speech. All recordings come from one microphone in one room, so there is no channel or environment variability. Emotional intensity was neither varied nor annotated, and no time-aligned or phonetic transcription is provided beyond the sentence identity. The cross-corpus comparisons reported in the accompanying article cannot be reproduced from this deposit, because EmoDB and RAVDESS are not ours to redistribute. Ethics. The study was approved by the Ethics Council of the University of Tirana, the institution implementing the project. All participants were adults who gave written bilingual informed consent, with separate opt-ins for recording, for public release and for demographic information, and who were informed that voice is identifying and that released recordings cannot be recalled from downloaded copies. No personally identifying information appears in any file or filename; speakers and evaluators appear only as anonymized identifiers. Only a blank consent template is distributed. Users are asked not to attempt re-identification. Carried out under a 2025 fellowship of the Research Expertise from the Academic Diaspora (READ) Program, funded by the Albanian-American Development Foundation (AADF) and administered in cooperation with the Institute of International Education (IIE).