Indonesian Emotional Speech from Counseling Sessions, University Students, and Theater Actors

Description

Indonesian emotional speech from three sources: counseling sessions, university students and professional theater actors, for use in speech emotion recognition research. Counseling (natural speech): 1,483 utterances from 44 respondents, cut by hand from counseling sessions recorded at Universitas Sumatera Utara and labeled with one of five emotions: neutral, happiness, sadness, fear or anger. Consent restrictions apply to the audio, so this subset is released as features. Each utterance is an MFCC matrix (NPY, 13 coefficients by n_frames), and counseling_mfcc.csv gives the mean and standard deviation of every coefficient. University Student (elicited speech): 325 utterances from four female students, who wrote their own sentences for the same five emotions. To lower the risk of identifying speakers by voice, this subset is also released as features. The MFCC matrices and the mean and standard deviation table follow the format of the counseling subset. Actor (acted speech): 840 WAV recordings from two professional theater actors, one female and one male, with 60 recordings per actor for each of seven emotions: the five above plus disgust and surprise. Each recording is a scenario of one or more sentences performed in the target emotion, so durations range from 2.7 to 73 s. MFCC matrices and a mean and standard deviation table are provided for these recordings in the same format as the other subsets, so all three subsets share one feature format. All audio was recorded in mono at 48 kHz, 32-bit float, with a wireless lavalier microphone. The WAV files are organized by actor and emotion code (01 neutral to 07 surprise). The metadata folder lists every file with its label, speaker, size, duration and SHA-256 checksum, along with statistics per class. README.md explains the folder layout, file names and variables.

Steps to reproduce

1. Recording: Audacity with an IT (Immersive Tech) wireless lavalier microphone, saved as WAV, mono, 48 kHz, 32-bit float. Counseling sessions were recorded in full. Students and actors were recorded one performance per file. 2. Counseling segmentation: respondent speech was cut by hand at natural pauses, ends of answers and topic changes. Counselor speech, introductions containing identifying information, very short responses and poor quality parts were removed. 3. Labeling: each counseling segment received one dominant emotion, based on listening and the counselor's session notes, and a psychologist reviewed the labels. Student and actor files carry the emotion the speaker was asked to perform, and one annotator checked each file. 4. Audio processing for student and actor recordings: silence at the start and end was trimmed, and noise reduction was applied. 5. Feature extraction for all three subsets: librosa MFCC with 13 coefficients, hop length 512 samples and the original 48 kHz sampling rate. Other settings are librosa defaults (n_fft 2048, Hann window, center=True). Each array is saved as float32 NPY without padding or truncation. Script: code/extract_mfcc.py. The actor features can be recomputed from the released WAV files. 6. Metadata: run code/compute_statistics.py <dataset_root> metadata [source_durations.csv] to regenerate file_list.csv, statistics_per_class.csv and label_mapping.csv. 7. Integrity: check downloaded files against CHECKSUMS.sha256, for example with shasum -a 256 -c CHECKSUMS.sha256.

Institutions

• Universitas Sumatera Utara

  North Sumatra, Medan

Categories

Licence