Access is granted only if you accept the terms below. In these terms, "EchoNova" means EchoNova AI L.L.C-FZ (Meydan Free Zone, Dubai, United Arab Emirates; licence 2650442.01). 1. You use the audio for research and evaluation only, under the terms on this card. 2. EchoNova makes no representation or warranty about the dataset. 3. You accept full responsibility for your use of the dataset and indemnify EchoNova against claims arising from it. 4. EchoNova may remove clips or end your access at any time.

EchoNova Arabic Dialect Benchmark (public subset)

مجموعة اختبار مفتوحة لقياس دقة تحويل الكلام العربي إلى نص: الفصحى وثماني لهجات، في 19.95 ساعة و8,925 مقطعًا، ولكل مقطع نص مرجعي كتبه إنسان.

A 20-hour subset of the private test set used in the EchoNova Sada technical paper: Modern Standard Arabic and eight dialects (Egyptian, Emirati, Saudi, Kuwaiti, Qatari, Omani, Syrian, Lebanese), cut into clips of at most 30 seconds, each with a human reference transcript. Hours per dialect are listed below. Headline figures in the paper are computed on the full 120-hour test set, of which this subset is one sixth; the remaining 100 hours are not released.

Results on this subset

Word error rate of the twelve systems compared in the paper, scored on this subset with the paper's normalization and scoring. Per-variety tables, confidence intervals, the scorer and the steps to reproduce these numbers are in the GitHub repository.

Word error rate of twelve systems on this subset. Sada-ASR is lowest at 11.74%.

System	Type	WER	Macro WER	CER	Empty %
Sada-ASR (EchoNova)	proprietary	11.74	15.56	4.72	0.00
Microsoft MAI-Transcribe-2	commercial API	13.79	17.07	6.52	0.01
Google Gemini 3.5 Transcribe	commercial API	16.31	20.80	8.39	0.03
ElevenLabs Scribe v2	commercial API	16.35	20.69	7.73	0.10
IFM Jais-ASR	open-weight	16.52	21.03	6.96	0.00
OpenAI gpt-transcribe	commercial API	17.11	21.72	8.61	0.01
Speechmatics Enhanced	commercial API	18.29	24.54	8.90	0.00
Audar-ASR-V1-Turbo	open-weight	21.01	25.29	10.09	0.00
NVIDIA FastConformer (Arabic)	open-weight	26.31	33.36	12.42	0.00
OpenAI Whisper large-v3	open-weight	26.82	32.72	15.17	0.00
NVIDIA Nemotron-3.5 ASR streaming 0.6B	open-weight	28.52	36.71	11.65	0.06
Cohere Transcribe Arabic	open-weight	29.53	38.52	19.75	0.03

Word error rate by variety and system on this subset

What is in the test set

Hours and clips per variety in this subset

Variety	العربية	Hours	Clips
MSA	الفصحى	8.27	3,400
Egyptian	المصرية	3.37	1,508
Emirati	الإماراتية	2.50	1,187
Saudi	السعودية	1.67	916
Kuwaiti	الكويتية	0.84	423
Qatari	القطرية	0.83	396
Lebanese	اللبنانية	0.83	363
Syrian	السورية	0.83	361
Omani	العُمانية	0.81	371
Total		19.95	8,925

Quick start

from datasets import load_dataset

# accept the access terms on this page and log in (huggingface-cli login) first
ds = load_dataset("echonova-ai/arabic-dialect-asr-benchmark", split="test")
clip = ds[0]  # clip["audio"]["array"] is 16 kHz mono; clip["text"] is the reference

Score your system's transcripts (one {"id": ..., "prediction": ...} JSON object per line) with the benchmark's scorer:

pip install "arabic-asr-benchmark[hf] @ git+https://github.com/echonova-ai-solutions/arabic-asr-benchmark"
asr-bench evaluate --predictions predictions.jsonl --system my-system --output my-system.json

Files

data/test-0000k-of-00004.parquet: the test split, one row per clip, with the columns id, audio (16 kHz, mono, FLAC), text (reference transcript), dialect, duration (seconds), recording_id (opaque source-recording label) and original_sampling_rate.

Scoring

Hypotheses and references are normalized as described in Section 5.2 of the paper. Raw outputs of the systems compared in the paper are not part of this release.

Terms of use

Source and ownership

The audio was collected from publicly available recordings on the internet. EchoNova does not own the copyright to the audio. Copyright remains with the original owners.

Licence

Transcripts, segmentation and metadata: CC BY 4.0. Audio: provided as found, for research and evaluation only. No commercial use of the audio, no redistribution of the audio outside this dataset.

Prohibited uses

Voice cloning, speaker identification or re-identification, and any attempt to identify, contact or profile the people recorded.

Access terms (gated)

In these terms, "EchoNova" means EchoNova AI L.L.C-FZ (Meydan Free Zone, Dubai, United Arab Emirates; licence 2650442.01).

You use the audio for research and evaluation only, under the terms above.
EchoNova makes no representation or warranty about the dataset.
You accept full responsibility for your use of the dataset and indemnify EchoNova against claims arising from it.
EchoNova may remove clips or end your access at any time.

Removal requests

Content owners or people recorded can request removal of any clip by emailing legal@echonova.ai with the clip ID. Clips are removed within 7 days and the dataset is re-versioned.

Citation

@misc{khaled2026sadaasr,
  title  = {Sada-ASR: A Production Multidialectal Arabic Speech Recognition System},
  author = {Khaled, Mahmoud and Abd-Elnaser, Mohammed and Mohsen, Mahmoud},
  year   = {2026},
  note   = {EchoNova technical report},
  howpublished = {\url{https://echonova.ai/research/pdf/EchoNova_Sada-ASR_Technical_Report.pdf}}
}

@misc{mohsen2026echonova,
  title        = {EchoNova Arabic Dialect ASR Benchmark},
  author       = {Mohsen, Mahmoud},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/echonova-ai/arabic-dialect-asr-benchmark}},
  note         = {EchoNova. Code and results: \url{https://github.com/echonova-ai-solutions/arabic-asr-benchmark}}
}

Downloads last month: 76

Total file size:

1.36 GB