---
license: cc-by-nc-4.0
language:
  - ar
task_categories:
  - automatic-speech-recognition
  - text-to-speech
  - audio-classification
tags:
  - speech
  - arabic
  - multi-dialect
  - emotion
  - emotional-speech
  - tts
  - asr
  - dialect-identification
  - najdi
  - hejazi
  - jordanian
  - moroccan
size_categories:
  - n<1K
pretty_name: "Arabic Multi-Dialect Emotional Speech — Demo (DataHive AI)"
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
---

# DataHive AI — Demo: Arabic Multi-Dialect Emotional Speech

A **DataHive AI** dataset: a stratified 1-hour demo sample from a full corpus of **50+ hours**. We can also create larger audio datasets upon client request.

Most public Arabic speech corpora flatten dialect into a single label and ignore emotion entirely. This corpus does the opposite: every recording is tagged with one of four regional Arabic dialects (Najdi, Hejazi, Jordanian, Moroccan) **and** one of four target emotions (Sad, Happy, Neutral, Angry), with two-reviewer QA on each utterance. The result is a dense, balanced grid useful for research and product work where regional voice authenticity and expressive prosody both matter.

## Quick load

```python
from datasets import load_dataset
ds = load_dataset("datahiveai/arabic-multidialect-emotional-speech-demo", split="train")
print(ds[0]["text"], ds[0]["dialect"], ds[0]["recording_style"], ds[0]["qa_tier"])
# Audio is decoded on access:
import soundfile as sf  # or any wav reader
sf.write("sample.wav", ds[0]["audio"]["array"], ds[0]["audio"]["sampling_rate"])
```

## Contact

- **Commercial licensing & enterprise inquiries**: [contact@datahive.ai](mailto:contact@datahive.ai)
- **Bug reports & data issues**: open a Discussion on the [HF dataset page](https://huggingface.co/datasets/datahiveai/arabic-multidialect-emotional-speech-demo/discussions)

## Use cases

- **Expressive Arabic TTS** — train or fine-tune a multi-speaker, multi-emotion
  voice model that handles Najdi / Hejazi / Jordanian / Moroccan natively
  instead of collapsing to MSA.
- **Emotion-conditional ASR & call-center analytics** — train models that not
  only transcribe but also tag utterances with speaker emotion (frustration
  detection, escalation triggers, sentiment dashboards).
- **Dialect identification** — build a 4-class classifier for routing,
  analytics, or upstream language-model conditioning in conversational AI.
- **IVR / voice-agent localization** — synthesize regional, emotionally
  appropriate voice replies for chatbots, banking IVR, healthcare prompts,
  and educational content targeted at specific Arabic-speaking markets.

## At a glance

- **Rows**: 496
- **Audio duration**: 62.3 minutes (~1.04 h)
- **Unique speakers**: 24
- **Audio**: 337 MB · 48 kHz · 16-bit · mono · WAV
- **Quality filter**: `qa_quality_tier ∈ {gold, silver}`

## Dialect × emotion grid

| | Angry | Happy | Neutral | Sad | **Total** |
|---|---:|---:|---:|---:|---:|
| **Hejazi** | 31 | 31 | 31 | 31 | **124** |
| **Jordanian** | 31 | 31 | 31 | 31 | **124** |
| **Moroccan** | 31 | 31 | 31 | 31 | **124** |
| **Najdi** | 31 | 31 | 31 | 31 | **124** |
| **Total** | **124** | **124** | **124** | **124** | **496** |

## Quality tier distribution

| Tier | Count |
|---|---:|
| gold | 396 |
| silver | 100 |

## Speaker gender

| Gender | Count |
|---|---:|
| female | 210 |
| male | 286 |

## QA scoring — what `qa_*` fields mean

Every recording in this demo is assessed by **two independent reviewers** on
a 6-dimension rubric (`qa_coverage = 2` for every row). We collapse the
rubric into a small set of derived fields that you can filter on directly.
All demo rows pass the gold / silver bar (`qa_quality_score ≥ 0.75`).

### Per-dimension value mapping

| Dimension | Mapping |
|---|---|
| Script Accuracy        | `perfect` = 1.00 · `acceptable` = 0.70 · `unusable` = 0.00 |
| Dialect / Accent Match | `yes` = 1.00 · `no` = 0.00 |
| Speech Quality         | `excellent` = 1.00 · `good` = 0.75 · `poor` = 0.25 · `unusable` = 0.00 |
| Background Noise       | `none` = 1.00 · `quiet` = 0.80 · `loud` = 0.20 |
| Technical Quality      | `excellent` = 1.00 · `good` = 0.75 · `low` = 0.40 · `poor` = 0.25 · `unusable` = 0.00 |
| Emotion match          | reviewer's perceived emotion == target `recording_style` → 1.00 else 0.00 |

### Weights (per-reviewer score)

`script_accuracy 0.20 · dialect_match 0.20 · speech_quality 0.20 ·
background_noise 0.15 · technical_quality 0.15 · emotion_match 0.10`. Blank
dimensions are dropped and the remaining weights renormalize.

The final `qa_quality_score` is the **mean** of the available reviewers' scores.

### Tier thresholds (on `qa_quality_score`)

| Tier | Range | In this demo |
|---|---|---|
| `gold`            | ≥ 0.90       | yes |
| `silver`          | 0.75 – 0.90  | yes |
| `bronze`          | 0.50 – 0.75  | excluded |
| `below_threshold` | < 0.50       | excluded |
| `unreviewed`      | no QA done   | excluded |

### Per-row QA fields

| Field | Meaning |
|---|---|
| `qa_coverage`                | Number of reviewers. |
| `qa_quality_score`           | 0–1 weighted score. |
| `qa_quality_tier`            | `gold` or `silver` in this demo. |
| `qa_emotion_match`           | `true` iff target emotion == reviewer's perception. |
| `qa_inter_reviewer_agreement`| Per-row agreement fraction between the two reviewers. |
| `qa_min_dimension`           | The lowest-scoring dimension on this row — useful for filtering. |

## Consent & provenance

Recordings were collected through DataHive AI's recording pipeline. Every
contributor in this corpus gave explicit, informed consent for their voice
to be recorded and used in the production of speech datasets.

All speaker identifiers in this dataset are opaque hashes — no real names,
phone numbers, addresses, or other directly-identifying personal information
is exposed.

## License

**This demo is licensed under [Creative Commons Attribution-NonCommercial
4.0 International (CC-BY-NC-4.0)](https://creativecommons.org/licenses/by-nc/4.0/)**
(SPDX: `CC-BY-NC-4.0`).

Two requirements: **(1) non-commercial use only, (2) attribution required**.

### What counts as non-commercial?

CC-BY-NC defines non-commercial as use **"not primarily intended for or
directed toward commercial advantage or monetary compensation."**
In practice for this demo:

**Allowed (non-commercial)**:
- Academic research, theses, papers, open peer-reviewed publications
- Educational use — university courses, tutorials, lectures, MOOCs that don't
  charge for access to materials trained on the demo
- Personal experiments, hobby projects, weekend hacks
- Public-good and non-profit work (NGO, government open initiatives,
  accessibility / assistive tools provided free of charge)
- Open-source ML benchmarks and research code, provided the project itself
  is non-commercial and free of charge
- Internal R&D / pre-product experimentation at any organization, as long as
  no model trained on this demo ships in a revenue-generating product

**Not allowed (commercial)**:
- Training a model that is sold, licensed, or served via a paid API or paid
  SaaS, even if the training step itself is internal
- Powering a feature inside a paid product (free tier of a paid SaaS counts
  as commercial — the "free tier" is a marketing channel for the paid one)
- Use in a freemium app where the free version drives users to a paid plan
- Inclusion in a paid course, textbook, conference workshop with a paid
  ticket, or any monetized media
- Training a model used in advertising, marketing automation, lead
  generation, or any revenue-attribution funnel
- Internal use at a for-profit company in any flow that ultimately
  contributes to a paid offering

If your use case sits in the gray zone (e.g., open-source project funded by
a commercial sponsor, free tier of a paid SaaS, ad-supported free app), the
default answer is **commercial** — get the commercial license.

### Required attribution

In **any** non-commercial use of this demo, the following credit must be
visible to anyone who could reasonably be expected to encounter the work
that uses it. Concretely:

| Where you use it | Where the credit must appear |
|---|---|
| Research paper / thesis | Citation in the references (use the BibTeX below) |
| Model card on Hugging Face / GitHub | "Training data" section + `datasets:` metadata field |
| Blog post / article / presentation slide | Footer, credit line, or attribution slide |
| Open-source code repository | `README.md` and / or `NOTICE` / `LICENSE-NOTICE.md` |
| Notebook shared publicly | Markdown cell at the top |

**Plain-text credit line** (for any of the above):

> Arabic Multi-Dialect Emotional Speech (Demo) — DataHive AI, 2026.
> Licensed under CC-BY-NC-4.0. https://datahive.ai

**BibTeX** (for academic citation):

```bibtex
@dataset{datahive_arabic_emotional_demo_2026,
  title     = {Arabic Multi-Dialect Emotional Speech (Demo)},
  author    = {DataHive AI},
  year      = {2026},
  note      = {1-hour stratified demo, 4 dialects × 4 emotions},
  license   = {CC-BY-NC-4.0},
  url       = {https://datahive.ai}
}
```

Removing or obscuring the credit, or replacing it with a generic "data
provider" placeholder, is a license violation.

### Want to use it commercially?

The full corpus is available under a custom commercial license that allows
training models for paid products, redistribution under your own terms, and
voice-cloning carve-outs. Contact your DataHive AI representative.

## What this demo is for

- Verifying audio quality, dialect authenticity, and emotion conveyance on
  real (not cherry-picked) samples.
- Smoke-testing your loading / training pipeline against the production schema.
- Internal evaluation prior to a full-corpus license from DataHive AI.

## What this demo is NOT

- Not a free-tier release. Use is restricted to internal evaluation under
  whatever evaluation terms accompanied this folder.
- Not a representative sample of the **unreviewed** partition (this demo is
  drawn entirely from reviewed rows).
- Not for redistribution.