About
Alconost·Speech dataset
Five Arabic dialects. Nine emotions.
Native speakers from five Arab regions perform everyday sentences in nine styles, from angry to whispered. Every recording is labelled and rated by native reviewers, and ready for training.
One sentence · nine ways
Real recordings from the dataset, compressed for the web. Press ▶ on any style to listen.
65,631recordings
70.1hof speech
61native voices
5dialects
9speaking styles
Made for
Voice products that speak your users' Arabic
Most Arabic speech data is formal and flat. Real users talk in dialect, and they get annoyed, excited and worried.
Expressive TTS
Voices that can sound cheerful, sad or urgent, with the right regional accent.
Emotion recognition
Spot frustration or distress in calls. Nine balanced, labelled classes.
Dialect ASR
Better recognition of colloquial, emotional and whispered speech.
Contact-centre AI
Nearly half the recordings come from healthcare, customer service and finance.
Dialect identification
Five dialects for routing and accent-aware models.
Benchmarking
Reviewer ratings of emotion and expression level on each recording help build targeted test sets.
Why this dataset
Built so models learn the emotion, not the speaker
6,254
one sentence, nine emotions
6,254 sentences are recorded in all nine styles within the same dialect. The words stay fixed and only the delivery changes, so a model learns the emotion itself and not the content around it. These sentences cover 87% of the dataset, and in 91% of them all nine readings come from the same speaker.
7
quality ratings on every recording
Native reviewers rate script accuracy, dialect match, speech quality, technical quality, background noise, perceived emotion and expression level. All scores ship with the data.
±1%
balanced emotion classes
The nine styles are evenly represented — 7,237 to 7,336 recordings each, all within 1% of the mean. No class reweighting before training.
EN ↔ AR
parallel text
Each line has an English source and a dialect-specific Arabic version, so the data also works for speech translation.
10
voices per dialect and style
Every dialect-and-style pair is recorded by ten native speakers for diversity.
61/61
cleared for AI training
Every speaker signed a release covering commercial use and AI/ML training, sublicensable, so the rights pass to you. Derivative work must not imitate an individual speaker. Full terms on request.
Coverage
Five dialects, evenly balanced
Hours per dialect, split by speaking style.
Share of recordings by topic.
Quality
Reviewed by native listeners
Every recording is rated by a native reviewer and the scores ship with the data, so you can set your own quality bar. Nothing is silently dropped: recordings the reviewers flagged as unusable stay in the dataset with their labels — use them as negative examples, or filter them out in one line.
Script accuracyperfect · acceptable · unusable
Dialect matchyes · no
Speech qualityexcellent · good · poor · unusable
Technical qualityexcellent · good · low · unusable
Background noisenone · quiet · loud
Emotionone of the 9 styles
Expression levelhigh · medium · low
63,799recordings pass QA (67.7 h)
1,832flagged unusable (2.8%), kept and labelled
60,063recordings with no quality flags (63.9 h)
97%with no loud background noise
81%where the reviewer heard the intended emotion
Two emotion labels on every recording
What the speaker was asked to convey, and what an independent native listener actually heard. They agree on 81% of recordings — train on the agreement subset for clean classes, or use the full set with perceptual labels.
Neutral94%
Whisper90%
Surprise85%
Sad82%
Angry80%
Happy77%
Fear75%
Disgust74%
Frustrated74%
Agreement between the target style and the emotion the reviewer perceived, across the recordings that pass QA.
In the box
Technical details
AudioWAV · 48 kHz · 16-bit · mono
RecordingsOne sentence each, 3.8 s on average
Text7,505 unique Arabic sentences, each with an English source
Speakers35 female, 26 male, aged 18–51
LabelsDialect, style, topic, gender, age
Review data7 reviewer ratings per recording
MetadataOne CSV row per recording, with anonymised speaker IDs
Localesar-EG · ar-SA (Hejazi, Najdi) · ar-JO · ar-MA
Need the full dataset? Get in touch.
Email us and we'll send it in the format you need — the complete set, or a slice: a single dialect, selected emotions, or only the recordings that pass QA. The full metadata schema and the licence terms come with it.
All 61 speakers signed releases covering commercial use and AI/ML training, and the rights pass to you.
QACAT
AI Video Studio