Persian Emotional Speech from Radio Plays

The Sharif Emotional Speech Dataset contains 3000 semi-natural utterances, equivalent to 3 hours and 25 minutes of speech data. It covers five basic emotions and a neutral state, labeled by 12 annotators from speech samples of 87 native-Persian speakers extracted from online radio plays.

Use Cases

• Train a speech emotion recognition model on 3000 utterances labeled for anger, fear, happiness, sadness, surprise, and neutral states.
• Analyze the distribution of emotional labels across 87 native-Persian speakers to study speaker-dependent emotional expression.
• Validate multi-annotator labeling consistency using the majority voting results from 12 annotators for each utterance.

Strengths

• 3000 speech utterances provide a substantial sample size for model training.
• Emotional labels are derived from majority voting among 12 annotators, potentially increasing reliability.
• Data is sourced from 87 different native-Persian speakers, offering speaker diversity.

Limitations

• The dataset is monolingual (Persian), limiting applicability to other languages.
• Emotional labels are limited to five basic categories and neutral, which may not capture complex emotional states.
• Data is extracted from semi-natural radio plays, which may not fully represent spontaneous, natural speech.

Provenance

Source

Online Persian radio plays.

Collection Method

Speech data extracted from radio plays and labeled by 12 annotators using majority voting.

Freshness

Last updated on 2022-10-24.

Geography

Persian language region.