Ramsa: A Large Sociolinguistically Rich Emirati Arabic Speech Corpus for ASR and TTS

The paper introduces Ramsa, a 41-hour sociolinguistically diverse Emirati Arabic speech corpus comprising 157 speakers across various subdialects and topics, which serves as a foundational resource for developing low-resource ASR and TTS technologies while establishing initial performance baselines for existing models.

The Gist

Imagine you are trying to teach a robot to speak and understand a specific local dialect of Arabic—the kind spoken in the United Arab Emirates (UAE). Until now, teaching this robot has been like trying to learn a language by reading a single, dusty page from a very old book. You might get the gist, but you'd miss the slang, the different accents, and the way real people actually talk to each other.

This paper introduces Ramsa, a massive new "library" of spoken Emirati Arabic designed to fix that problem. Here is a simple breakdown of what the researchers did, using some everyday analogies.

1. The Problem: The "Empty Bookshelf"

For a long time, AI researchers had plenty of data for English or standard Arabic, but very little for the specific, colorful dialects of the Gulf region.

• The Old Data: Previous collections were like a small, one-room library. Some had only one speaker (usually a man), some were too short, and they mostly ignored the different "accents" within the UAE (like the city voice vs. the desert/Bedouin voice vs. the mountain voice).
• The Gender Gap: Many old libraries were almost entirely male. If you only listen to men, you don't learn how women speak, which is a huge missing piece of the puzzle.

2. The Solution: Building "Ramsa" (The New Library)

The researchers built Ramsa, a 41-hour collection of audio recordings. Think of this as building a massive, modern museum of sound.

• Who is in the museum? They recorded 157 different people (59 women and 98 men). This is a huge improvement over previous collections that might have had only one or two women.
• What are they talking about? They didn't just ask people to read a script. They recorded:
  • Structured Interviews: Like a friendly chat over coffee about daily life, food, and family.
  • TV Shows: Clips from real Emirati television, featuring everything from cooking shows and history documentaries to talk shows about business and culture.
• The Variety: The library captures the full spectrum of the UAE. You can hear the "Urban" city accent, the "Bedouin" desert accent, and the "Mountain" accent. This is crucial because AI needs to understand that a word might sound slightly different depending on where the speaker grew up.

3. The "Translation" Challenge (Annotation)

Once they had the audio, they had to write it down (transcribe it). This is tricky because Emirati Arabic is spoken, not just written.

• The Rule: They decided to write down exactly how the words sound, not how they are spelled in a formal dictionary.
  • Example: If someone says "shay" (something) instead of the formal "shayʾ", they wrote "shay." If words get mashed together because people speak fast, they wrote them mashed together.
• The Team: A team of linguists acted as the "translators," carefully listening to every second to ensure the written text matched the sound perfectly, including laughter, interruptions, and pauses.

4. The Test Drive: Can AI Understand It?

The researchers took a small slice (10%) of this new library and tested it against the world's best AI speech tools (both free and paid) to see how well they could understand Emirati Arabic without any special training first (this is called a "zero-shot" test).

• The Results:
  • The Winners: The open-source model Whisper-large-v3-turbo did the best job at understanding speech (ASR). For turning text into speech (TTS), MMS-TTS-Ara was the champion.
  • The Struggles: The AI struggled the most with fast, overlapping conversations (like a chaotic cooking show where people talk over each other). It was much better at listening to a single person telling a story (like a documentary).
  • The Verdict: The AI is getting there, but it's not perfect yet. It's like a student who passed the test but still needs to study more to get an A+.

5. Why This Matters (The Big Picture)

This paper isn't just about numbers; it's about representation.

• For Technology: It gives AI developers the "training wheels" they need to build better voice assistants, dictation tools, and translation apps specifically for Emiratis.
• For Culture: It preserves the way Emiratis actually speak today, capturing the mix of city and desert accents that are blending together.
• For the Future: The researchers admit this is just the beginning. They need more data from the mountain regions and more older speakers to capture how the language changes over time.

In a nutshell: The authors built a giant, diverse sound library of Emirati Arabic to teach AI how to really listen and speak like a local. They tested the top AI models, found they are good but still need practice, and opened the door for future researchers to build even smarter tools for the UAE.

Technical Summary

Here is a detailed technical summary of the paper "Ramsa: A Large Sociolinguistically Rich Emirati Arabic Speech Corpus for ASR and TTS."

1. Problem Statement

Despite the increasing demand for dialect-aware speech technologies, Gulf Arabic, and specifically Emirati Arabic, remains underrepresented in large-scale speech corpora. Existing resources suffer from three primary limitations:

• Scale and Metadata: Many corpora are small (e.g., 4–14 hours) or lack critical sociolinguistic metadata such as gender distribution and subdialect classification.
• Gender Imbalance: Existing datasets often skew heavily toward male speakers (e.g., Alsanaa has only one male; Casablanca is only ~25% female).
• Homogeneity: Most resources treat Emirati Arabic as a monolithic variety, ignoring well-documented internal variations across Urban, Bedouin, and Mountain/Shihhi subdialects.
• Lack of Baselines: There are no established zero-shot benchmarks for Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) specifically for Emirati Arabic, hindering the development of low-resource language technologies.

2. Methodology

Corpus Construction (Ramsa)

The authors developed Ramsa, a 41-hour speech corpus designed to address the gaps above. Data was sourced from two distinct channels:

1. Structured Interviews: 14 native speakers (13 female, 1 male) recruited from the University of Sharjah. Participants were categorized by gender, age, birth emirate, and self-identified subdialect (Urban, Bedouin, Mountain/Shihhi, or Mixed). Recordings were monologic, conducted in quiet office settings.
2. Television Broadcasts: Ten programs from national Emirati channels (e.g., Emarat TV, Sharjah TV) covering diverse topics (heritage, agriculture, architecture, daily life). These provided both monologic and dialogic formats, varying acoustic conditions (background music, noise), and speakers from different regions (Abu Dhabi, Sharjah, Liwa Oasis, Al Dhaid).

Sociolinguistic Diversity:

• Speakers: 157 total speakers (59 female, 98 male).
• Subdialects: Includes Urban, Bedouin, Mountain/Shihhi, and Mixed categories.
• Formats: 91 monologic and 79 dialogic recordings with varying lengths (<5 min to >15 min).

Annotation Strategy

• Transcription: Four linguistically trained annotators (3 native Emirati speakers, 1 non-native) performed the work.
• Orthography: The corpus adopts a dialectal orthographic strategy rather than normalizing to Modern Standard Arabic (MSA).
  • Phonological Reductions: Retained as produced (e.g., mā shāʾ Allāh $\to$ mashāllah).
  • Dialect-Specific Substitutions: Transcribed phonetically (e.g., /q/ $\to$ /g/, /k/ $\to$ /sh/).
  • Variation: Multiple realizations of the same word are transcribed exactly as spoken.
• Segmentation: Speech is segmented into "utterances" defined as complete communicative acts (statements, questions, requests) based on prosodic and interactional cues.

Benchmarking Protocol

A 10% subset (approx. 4 hours, 29,301 tokens) was manually transcribed to establish zero-shot baselines.

• ASR Evaluation: Tested against three models:
  • AssemblyAI Universal-2 (Commercial, Conformer–RNN-T).
  • Gladia AI (Solaria/Whisper-Zero) (Commercial, Whisper-based).
  • Whisper-large-v3-turbo (Open-source, Transformer).
  • Metric: Word Error Rate (WER) and Character Error Rate (CER) at the document level.
• TTS Evaluation: Tested against two open-source models:
  • ArTST (SpeechT5 framework).
  • MMS-TTS-Ara (Meta's Massively Multilingual Speech).
  • Metric: Back-transcription using Whisper-large-v3-turbo to calculate WER/CER, ensuring evaluation of intelligibility independent of the TTS model's internal architecture.

3. Key Contributions

1. Ramsa Corpus: The introduction of the largest and most sociolinguistically diverse Emirati Arabic corpus to date (41 hours), featuring improved female representation (37.5%) and explicit coverage of subdialectal variation.
2. Annotation Guidelines: A comprehensive framework for transcribing Emirati Arabic that preserves dialectal phonology and orthography without normalizing to MSA, crucial for low-resource dialect modeling.
3. Zero-Shot Baselines: The first established performance benchmarks for Emirati Arabic ASR and TTS using state-of-the-art commercial and open-source models in a zero-shot setting.

4. Results

ASR Performance

• Best Model: Whisper-large-v3-turbo achieved the best average performance with a WER of 0.268 and CER of 0.144.
• Comparison: Outperformed AssemblyAI (WER 0.354) and Gladia (WER 0.347).
• Source Variance:
  • Easiest: Monologic interviews and documentaries (e.g., Qiṣaṣmin Liwa).
  • Hardest: Overlap-rich conversational formats (e.g., Maṭbakh al-Dār), where WER spiked to ~0.73–0.79 due to rapid turn-taking and interruptions.
• Context: Results are significantly better than previous benchmarks on the Casablanca corpus (Whisper-large-v2 WER 0.52).

TTS Performance

• Best Model: MMS-TTS-Ara outperformed ArTST with an average WER of 0.285 and CER of 0.081.
• Observations:
  • MMS-TTS-Ara showed superior performance in interviews and informational speech.
  • Both models struggled with overlap-dense formats (e.g., Maṭbakh al-Dār WER ~0.45–0.58).
  • CER was consistently lower than WER, indicating that character-level metrics are more robust for dialectal orthography where tokenization is challenging.

5. Significance and Future Outlook

• Foundation for Low-Resource Tech: Ramsa provides the necessary data foundation to fine-tune models for Emirati Arabic, moving beyond the limitations of zero-shot generalization.
• Sociolinguistic Utility: The corpus enables research into dialectal variation, language change, and gender dynamics in speech, which were previously obscured by homogeneous datasets.
• Challenges Identified:
  • Subdialect Imbalance: While improved, Bedouin and Mountain/Shihhi speakers remain underrepresented compared to Urban speakers.
  • Generational Gaps: Annotators (aged 20–40) reported difficulty with certain lexical items, suggesting language attrition or change that requires broader age-range involvement in future annotation.
  • Evaluation Limits: Current benchmarks are zero-shot; future work must focus on fine-tuning and domain adaptation to reach the upper bounds of performance.
• Data Access: The interview data is available under restricted access for non-commercial research (via Data Use Agreement), while broadcast data is copyright-restricted but supported by metadata and source URLs for reproducibility.

In summary, Ramsa represents a critical step forward in democratizing Emirati Arabic speech technology, offering a rich, annotated resource that bridges the gap between sociolinguistic research and computational modeling.