Sada-ASR: A Production Multidialectal Arabic Speech Recognition System

Mahmoud Khaled, Mohammed Abd-Elnaser, Mahmoud Mohsen, Hazem Ahmed, Abdelrahman HafezarXiv 2026 · arXiv

TL;DR

Sada-ASR is a production-grade speech recognition system designed to handle the "diglossic" nature of Arabic, where formal written language (MSA) differs vastly from spoken regional dialects. By leveraging a high-quality human-in-the-loop data pipeline, it outperforms major commercial offerings from Microsoft, OpenAI, and Google across a diverse 120-hour multi-dialect benchmark.

What Problem Does This Paper Address?

The core challenge in Arabic Automatic Speech Recognition (ASR) is diglossia. While Modern Standard Arabic (MSA) is the standard for news and formal writing, it is rarely spoken in daily life or customer service interactions. Instead, people use regional dialects (e.g., Egyptian, Gulf, Levantine) that vary significantly in phonology, vocabulary, and grammar. Furthermore, dialects lack a standardized orthography (spelling rules) and frequently incorporate English code-switching.

Most existing ASR systems are trained primarily on MSA. When deployed in production environments like voice agents, these systems often "mishear" the dialect, causing the interaction to fail before the underlying logic (like an LLM) even has a chance to process the request. The authors aim to build a system that is robust enough for production-level customer service across the diverse Arab world.

Methods

Sada-ASR utilizes an attention-based encoder-decoder architecture that generates transcripts token-by-token. While the architecture itself is relatively standard for modern ASR, the "secret sauce" lies in their data pipeline and training methodology:

  • Two-Stream Data Pipeline: They combine high-volume, "weakly labeled" web data with high-quality transcripts from an in-house platform featuring over 100 annotators, reviewers, and domain experts.
  • Rigorous Quality Control: To ensure consistency across dialects, they use "control samples"—the same audio clips inserted into different annotators' queues—to measure and normalize performance.
  • Balanced Training: Each training epoch is split 50/50 between Arabic and English to handle code-switching. Within the Arabic portion, clips are sampled in proportion to the hours available for each dialect.
  • Production-Ready Decoding: Instead of complex beam searches, they use a "plain greedy search" with specific guards to prevent repetition loops and cap output length, ensuring stability and low latency for live calls.

Datasets

The paper evaluates Sada-ASR on a mix of private and public benchmarks:

  • Private Multi-Dialect Test Set: A new 120-hour dataset covering MSA and eight dialects (Egyptian, Saudi, Syrian, Emirati, Kuwaiti, Qatari, Omani, and Lebanese). The authors have released a 20-hour subset of this to the community.
  • Public Benchmarks: Evaluation on six established sets, including SADA (Saudi), Common Voice, MASC, MGB-2 (Broadcast), and Casablanca (Moroccan).
  • Training Data: A massive corpus including thousands of hours of web audio and a significant portion of English speech (about one-third of total hours).

Key Insights and Ideas

  • The "MSA-Normalization" Penalty: One of the most interesting findings is why the system "lost" on the Lebanese dialect. Sada-ASR tended to write Lebanese words in their MSA forms (e.g., writing kathīr instead of the spoken ktīr). Because the ground-truth references were verbatim, the system was penalized for "correcting" the dialect into formal Arabic.
  • Data Quality > Augmentation: Surprisingly, the authors achieved state-of-the-art results without common data augmentation techniques like SpecAugment, noise injection, or speed perturbation. This suggests that for complex linguistic tasks like dialectal Arabic, the quality and diversity of human-labeled data are more critical than synthetic variety.
  • The Importance of Per-Variety Metrics: Aggregate Word Error Rate (WER) can be misleading. A system can look great on average by being perfect at MSA but failing at Omani. By breaking down results by dialect, the authors highlight that Sada-ASR’s biggest wins are in the most challenging regional varieties.

Key Results

Sada-ASR demonstrated superior performance against eleven major competitors, including Microsoft MAI-Transcribe-2, OpenAI Whisper large-v3, and Google Gemini 3.5:

  • Overall Performance: Achieved an 11.69% WER on the private test set, a 13.5% relative reduction over the strongest external system (Microsoft at 13.51%).
  • Dialectal Wins: It showed massive leads in specific dialects, such as Omani (4.30 point lead) and Egyptian (3.78 point lead).
  • Public Leaderboards: On six public test sets, it averaged a 17.17% WER, outperforming published figures for specialized models like Falcon-ASR (20.92%) and Jais-ASR (22.85%).
  • Reliability: The system had a near-zero "empty transcript" rate (0.00%), whereas some commercial APIs failed to return any text for up to 0.10% of clips.

Limitations and Future Work

  • Geographic Coverage: The current test set lacks representation for Maghrebi (North African), Iraqi, Sudanese, and Yemeni Arabic.
  • Verbatim vs. Formal: The authors acknowledge the need to train on more "verbatim" Levantine transcripts to avoid the penalty of auto-correcting dialects into MSA forms.
  • End-to-End Evaluation: While the ASR is strong, the authors plan to conduct future "listening tests" of the full voice agent (ASR + LLM + TTS) to measure user satisfaction and turn latency in real-world scenarios.

Summary updated

Abstract

Callers to Arabic customer-service lines speak their own dialect, while speech recognizers are built and evaluated mainly on Modern Standard Arabic (MSA). A recognizer that mishears the dialect breaks the call before anything else can help. We present Sada-ASR, the speech recognizer of EchoNova's production Arabic voice agent. On a 120-hour test set of MSA and eight dialects, it has the lowest word and character error rates of twelve systems, against eleven from Microsoft, ElevenLabs, Google, OpenAI, Speechmatics, IFM, Cohere, NVIDIA and Audar: 11.69% word error rate against 13.51% for the strongest, Microsoft's MAI-Transcribe-2 (a lead of 0.74 points on clips unaffected by that system's fixed number and script conventions), and 15.52% against 16.96% averaged over varieties. Its largest margins are on Omani (4.30 points), Egyptian (3.78) and Saudi (2.19) speech; it is level on Syrian and loses on Lebanese (24.96% against 23.05%), mostly because it writes Lebanese words in MSA forms and is less verbatim. On six public test sets its average word error rate under the leaderboard's current scoring is 17.17%, below every published figure we found, among them Falcon-ASR (20.92%), Jais-ASR (22.85%) and Audar-ASR-V1-Turbo (23.17%), though none of these states that it used that scoring, so the comparison is indicative. We describe the data pipeline and methodology behind Sada-ASR and release a 20-hour subset of the test set with our scoring code.
BibTeX
@misc{khaled2026sada,
  author = {Mahmoud Khaled and Mohammed Abd-Elnaser and Mahmoud Mohsen and Hazem Ahmed and Abdelrahman Hafez},
  title = {{Sada-ASR: A Production Multidialectal Arabic Speech Recognition System}},
  year = {2026},
  eprint = {2611.19512},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19512},
}
Figure