Research · Technical report · October 2026

The most accurate model for Arabic dialects.

We tested Sada-ASR, the speech recognition inside EchoNova, against eleven systems from Microsoft, ElevenLabs, Google and others, on the same 120 hours of real Arabic speech. It made the fewest mistakes.

Read the paper PDF · 10 pages Test set Hugging Face Scoring code GitHub

Published October 2026Systems tested 12Audio 120 hours · 53,597 clipsVarieties Fusha + 8 dialects

The result

Fewer wrong words than any system we tested.

Word error rate is the share of words a system gets wrong. Every system heard the same audio and was scored the same way.

Word error rateEach dot is one wrong word in 100 · lower is better

EchoNovaSada-ASR

11.69%

MicrosoftMAI-Transcribe-2

13.51%

ElevenLabsScribe v2

16.12%

GoogleGemini 3.5 Transcribe

16.13%

IFMJais-ASR

16.41%

SpeechmaticsEnhanced

18.08%

AudarASR V1 Turbo

20.86%

OpenAIWhisper large-v3

26.57%

Measured by EchoNova, 13 to 16 October 2026 · All 12 systems and per-dialect tables in the paper

In plain words

What we did, and how to read it.

What we measured

How often each system writes a word wrong. Sada-ASR's 11.69% means about 12 wrong words in every 100. The next best system, Microsoft's, gets about 14 wrong.

How we kept it fair

Every system heard the same 53,597 clips, was scored by the same code and was run by us in the same week. None of the test audio was used to train Sada-ASR.

Where we lost

On Lebanese Arabic, Microsoft makes fewer word errors than we do, and on Syrian we are level. We show it because a test we never lose is not worth trusting.

By dialect

First on 7 of 9 Arabic varieties.

Sada-ASR's word error rate in each variety, next to the strongest other system.

Fusha1st

5.61%

next best Microsoft 6.99%

Egyptian1st

9.48%

next best Microsoft 13.26%

Saudi1st

14.74%

next best Microsoft 16.93%

Emirati1st

18.57%

next best Microsoft 20.22%

Kuwaiti1st

14.29%

next best Microsoft 14.83%

Qatari1st

15.82%

next best Microsoft 17.02%

Omani1st

16.07%

next best Microsoft 20.37%

SyrianLevel

20.16%

level with Microsoft 19.98%

Lebanese2nd

24.96%

behind Microsoft 23.05%

Every dialect, every system

All ten systems we ran, in every variety. Darker cells mean more errors. Rows are ordered by overall word error rate.

	All	Fusha	Egyptian	Saudi	Emirati	Kuwaiti	Qatari	Omani	Syrian	Lebanese
EchoNovaSada-ASR	11.69	5.61	9.48	14.74	18.57	14.29	15.82	16.07	20.16	24.96
MicrosoftMAI-Transcribe-2	13.51	6.99	13.26	16.93	20.22	14.83	17.02	20.37	19.98	23.05
ElevenLabsScribe v2	16.12	7.80	15.19	21.36	24.40	18.24	21.72	23.05	24.90	29.37
GoogleGemini 3.5 Transcribe	16.13	8.03	14.32	20.68	24.85	19.21	20.73	22.35	25.70	31.73
IFMJais-ASR	16.41	7.83	16.55	21.06	23.86	19.84	20.32	22.65	26.54	30.35
SpeechmaticsEnhanced	18.08	7.82	17.56	20.94	26.78	33.84	22.32	22.81	32.89	34.37
AudarASR V1 Turbo	20.86	11.04	25.45	23.95	26.86	23.22	22.96	26.45	31.63	34.93
NVIDIAFastConformer	25.97	12.05	28.65	31.72	37.22	35.48	35.33	34.55	38.55	42.84
OpenAIWhisper large-v3	26.57	14.66	28.11	31.11	37.24	31.02	31.95	33.93	39.60	45.41
NVIDIANemotron 3.5	28.11	12.35	31.57	34.40	37.95	35.71	33.89	36.23	49.74	55.32

	All	Fusha	Egyptian	Saudi	Emirati	Kuwaiti	Qatari	Omani	Syrian	Lebanese
EchoNovaSada-ASR	4.76	2.08	3.73	5.61	8.54	4.79	7.56	7.77	7.79	10.86
MicrosoftMAI-Transcribe-2	6.37	3.58	6.56	7.53	9.44	5.43	8.56	10.02	7.81	11.76
ElevenLabsScribe v2	7.54	3.72	7.33	9.72	11.60	6.69	10.85	11.55	10.43	15.52
GoogleGemini 3.5 Transcribe	8.23	4.17	7.31	9.79	13.56	8.03	11.48	11.96	11.92	18.20
IFMJais-ASR	6.89	3.21	6.67	8.57	10.83	7.20	9.65	11.15	10.64	13.82
SpeechmaticsEnhanced	8.73	3.57	8.17	9.79	13.97	17.56	11.76	11.19	15.51	18.35
AudarASR V1 Turbo	9.97	6.22	11.62	10.17	13.24	9.11	11.53	13.57	13.64	16.88
NVIDIAFastConformer	12.31	6.33	11.92	14.24	18.83	15.91	18.74	18.35	17.53	22.01
OpenAIWhisper large-v3	15.08	9.68	14.49	16.30	22.02	15.46	18.38	20.37	21.95	25.79
NVIDIANemotron 3.5	11.43	4.72	12.49	13.52	16.73	12.72	15.14	17.22	20.84	24.81

Best system in that dialectLevel, inside the margin of errorFewer errorsMore errors

Lebanese is our one loss, 1.91 points behind Microsoft. Syrian is a tie. By character error rate, Sada-ASR is lowest in all nine varieties. Full tables, including systems not shown here, are in the paper.

How we built it

People who speak the dialect, checking every word.

Dialects have no standard spelling and few written transcripts. So we built our own annotation platform, with native speakers from across the Arab world at every step.

100+ annotators

In-house native speakers transcribe the audio, word for word, in their own dialect.

Reviewers

A second layer reviews and audits every submission before it counts.

Experts

Professional writers and domain specialists settle the hardest cases.

Hidden checks

Control and duplicate samples measure every annotator, then are removed before training.

Public test sets

Also first on the six public Arabic test sets.

The Open Universal Arabic ASR Leaderboard uses six public test sets. We scored Sada-ASR on them with the leaderboard's own code. Other rows are figures their makers published, so the comparison is indicative.

System	Average word error rate	Source
EchoNova Sada-ASRMeasured by EchoNova	17.17%	This report
TII Falcon-ASR	20.92%	Published by TII
IFM Jais-ASR	22.85%	Published by IFM
Audar-ASR V1 Turbo	23.17%	Leaderboard
Speechmatics Melia 1	23.51%	Measured by IFM
ElevenLabs Scribe v2	24.54%	Measured by IFM

Method

For engineers.

The test set

120 hours in 53,597 clips of at most 30 seconds, cut from 14,686 source recordings: Fusha (Modern Standard Arabic) and eight dialects, Egyptian, Emirati, Saudi, Kuwaiti, Qatari, Omani, Syrian and Lebanese. Each variety contributes a fixed number of hours. No clip of this set, and no part of the six public test splits, was used to train or select Sada-ASR.

Scoring

Word and character error rates are corpus micro-averages: total edit distance divided by total reference length. Hypotheses and references are normalized identically: punctuation and diacritics removed, hamza and madda forms of alif folded, Eastern Arabic digits mapped to Western, word-final taa marbuta folded into haa. Numbers are not converted between digits and words.

Confidence

Margins carry 95% confidence intervals from 2,000 bootstrap resamples of source recordings. Sada-ASR's lead over Microsoft MAI-Transcribe-2 is 1.83 points (95% CI 1.72 to 1.93). The paper also reports the margin on clips where Microsoft's fixed number and script conventions do not apply.

Systems and settings

Every API was set to Arabic and to verbatim output where possible. No system, ours included, used contextual biasing, custom vocabulary or any other adaptation. Every run is dated, and raw outputs are stored so every figure can be recomputed without re-running a recognizer.

Model

Sada-ASR is an attention-based encoder-decoder: a Conformer encoder turns audio into acoustic representations and a Transformer decoder writes the transcript. It is decoded with plain greedy search on every test set, with no decode-time tuning.

Check it yourself

Everything you need to rerun it.

Paper · PDFSada-ASR technical reportData pipeline, method, full results for 12 systems and the error analysis.Download →Dataset · Hugging Face20-hour public test setThe public release: 8,925 clips across all nine varieties with reference transcripts. Results on this page use the full 120-hour set. Research use, gated access.Open →Code · GitHubScoring and normalizationThe exact code behind every number on this page. Apache 2.0.Open →

@techreport{echonova2026sada,
  title       = {Sada-ASR: A Production Multidialectal Arabic Speech Recognition System},
  author      = {Khaled, Mahmoud and Abd-Elnaser, Mohammed and Mohsen, Mahmoud},
  institution = {EchoNova, United Arab Emirates},
  year        = {2026},
  month       = {oct}
}

For media

The facts, ready to quote.

Sada-ASR made the fewest word errors of 12 speech recognition systems tested on Arabic.
It got about 12 words wrong in 100, against about 14 for the next best system.
It came first on 7 of 9 Arabic varieties, from Fusha to Gulf, Egyptian and Levantine dialects.
Built in the UAE, trained on speech transcribed by more than 100 native speakers.

Press enquiries

Questions

What is word error rate?

The share of words a system writes wrong: swapped, missed or added. Lower is better. 10% means one wrong word in ten.

Why are Arabic dialects hard?

People speak dialect, but most written Arabic is Fusha. Dialects have no standard spelling, mix in English and differ from country to country, so systems trained mainly on Fusha miss words people actually say.

Who checked these results?

We ran every system ourselves, dated every run and published the test set and scoring code so anyone can rerun them. Independent reproductions are welcome.

Can I try it on my own audio?

Yes. Sign up and use Speech to Text in the playground, or talk to us about a pilot.