datahiveai

whisper-large-v3-ar-dialects

Results (held-out target test · n = 3,354 · speaker/script/text-disjoint)

Lower is better. WER / CER under the Open Universal Arabic ASR Leaderboard normalizer.

Table with columns: Dialect, Stock Whisper-large-v3, + our corpus (this adapter)
Dialect	Stock Whisper-large-v3	+ our corpus (this adapter)
Najdi (MSA-adjacent)	0.370 / 0.103	0.162 / 0.043
Hejazi	0.388 / 0.115	0.194 / 0.055
Jordanian	0.512 / 0.206	0.248 / 0.077
Moroccan Darija	0.804 / 0.337	0.342 / 0.105
Overall	0.589 / 0.226	0.258 / 0.077

Versus strong baselines (overall, same test):

Table with columns: WER, CER
	WER	CER
Stock Whisper-large-v3	0.589	0.226
Public-only fine-tune (SADA + MASC)	0.598	0.204
Meta omniASR-LLM-7B (open, 7B)	0.455	0.140
Deepgram nova-3 (commercial API)	0.493	0.137
+ our corpus (this adapter)	0.258	0.077

We report CER alongside WER because dialectal Arabic has no standard orthography and clitics agglutinate onto words, which makes word-level WER swing with spelling; CER is the more trustworthy signal. All numbers come from a single consistent scoring pass with 95% bootstrap CIs; see the whitepaper for the convention-robust (CAMeL) re-scoring and full deltas.

Model description

Developed by: DataHive AI
Model type: LoRA (PEFT) adapter for automatic speech recognition
Base model: openai/whisper-large-v3
Language: Arabic (dialectal — Hejazi, Najdi, Jordanian, Moroccan Darija)
Task: transcribe (Arabic)
Access terms: gated — granted for evaluation; please don't redistribute the weights (agreed to at request time). No separate license file; the access request is the agreement.

This adapter is fine-tuned on a mixture of public Arabic ASR data (SADA, MASC) plus DataHive AI's proprietary 4-dialect × 4-emotion corpus — native-speaker, conversational recordings across ten everyday domains, collected on DataHive AI's data-collection platform. A public demo subset of that corpus is available at datahiveai/arabic-multidialect-emotional-speech-demo; the full corpus is proprietary and not released. This card documents the model it produces.

A controlled ablation isolates the corpus as the cause of the lift: training on the public data alone leaves error essentially at stock; adding our corpus (holding the public mixture identical) is what moves the needle.

Intended uses & limitations

Direct use. Transcription of conversational, dialectal Arabic in the covered dialects (Hejazi, Najdi, Jordanian, Moroccan Darija), including emotional/expressive speech.

Out-of-scope / use with care.

Not a universally better Arabic model. The advantage is specialization for this domain and these dialects/convention. On neutral, out-of-domain benchmarks (e.g. MSA-adjacent Saudi speech in a third-party convention) frontier systems are competitive, and on some external Moroccan sets a larger open model is ahead. See the whitepaper for the honest, non-cherry-picked comparison.
Dialects and domains outside the training coverage are not guaranteed; results reflect our dialectal convention.
ASR output can contain errors; do not use for high-stakes decisions without human review.

How to get started

The weights are gated — request access on this page first. Once granted:

python
from peft import PeftModel
from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch, torchaudio

BASE = "openai/whisper-large-v3"
ADAPTER = "datahiveai/whisper-large-v3-ar-dialects"

processor = WhisperProcessor.from_pretrained(BASE)
model = WhisperForConditionalGeneration.from_pretrained(BASE, torch_dtype=torch.float16)
model = PeftModel.from_pretrained(model, ADAPTER)   # requires access token with granted access
model = model.merge_and_unload().to("cuda").eval()

wav, sr = torchaudio.load("clip.wav")
if sr != 16000:
    wav = torchaudio.functional.resample(wav, sr, 16000)
feats = processor(wav.squeeze().numpy(), sampling_rate=16000, return_tensors="pt").input_features
feats = feats.to("cuda", torch.float16)

ids = model.generate(feats, language="ar", task="transcribe")
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

Training details

Base: openai/whisper-large-v3 (LoRA / PEFT — only the adapter is trained)
LoRA: r = 64, α = 128, dropout = 0.05, target modules = q_proj, k_proj, v_proj, out_proj
Optimizer / schedule: lr 1e-4, warmup 500 steps, weight decay 0.01
Batch: 8 × grad-accum 2 (effective 16), seed 42, fp16
Data mixture: 40% proprietary corpus, 33.75% SADA, 26.25% MASC (public part held byte-identical between the public-only and public+ours ablation recipes)
Audio augmentation: noise/RIR augmentation applied more heavily to the proprietary corpus than to the already-in-the-wild public data

Evaluation

Test set: proprietary held-out, speaker + script + text-disjoint (~3,354 clips, four dialects) — no speaker, script, or sentence is shared with training, so results reflect genuine generalization.
Metrics: WER and CER, under two normalizers — the Open Universal Arabic ASR Leaderboard normalizer (faithful) and a convention-robust CAMeL-standard normalizer; 95% CIs from 2,000× paired bootstrap.
Honest caveat: part of any model's apparent lead on matched data is orthographic convention; under convention-robust scoring our lead narrows but holds against strong baselines, and is essentially unchanged versus stock (real quality, not spelling). Full tables, external transfer tests, and the losses we don't win are in the whitepaper.

Citation

bibtex
@misc{datahive_whisper_ar_dialects,
  title  = {The Data Moat in Dialectal Arabic Speech Recognition},
  author = {DataHive AI},
  year   = {2026},
  url    = {https://datahive.ai/blog/2026/07/06/the-data-moat-in-dialectal-arabic-speech-recognition/},
  note   = {Whisper-large-v3 + LoRA on a proprietary 4-dialect Arabic corpus}
}

Data & licensing — get in touch

The lift here comes from targeted data. If you need dialectal/emotional Arabic ASR for your own domain — or a targeted corpus collected to your spec — DataHive AI builds the data and proves the lift on your test set. Contact: contact@datahive.ai.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

datahiveai

Model Tree

Base

openai/whisper-large-v3

Adapter

this model

Input Modalities

Audio

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Results (held-out target test · n = 3,354 · speaker/script/text-disjoint)

Lower is better. WER / CER under the Open Universal Arabic ASR Leaderboard normalizer.

Table with columns: Dialect, Stock Whisper-large-v3, + our corpus (this adapter)
Dialect	Stock Whisper-large-v3	+ our corpus (this adapter)
Najdi (MSA-adjacent)	0.370 / 0.103	0.162 / 0.043
Hejazi	0.388 / 0.115	0.194 / 0.055
Jordanian	0.512 / 0.206	0.248 / 0.077
Moroccan Darija	0.804 / 0.337	0.342 / 0.105
Overall	0.589 / 0.226	0.258 / 0.077

Versus strong baselines (overall, same test):

Table with columns: WER, CER
	WER	CER
Stock Whisper-large-v3	0.589	0.226
Public-only fine-tune (SADA + MASC)	0.598	0.204
Meta omniASR-LLM-7B (open, 7B)	0.455	0.140
Deepgram nova-3 (commercial API)	0.493	0.137
+ our corpus (this adapter)	0.258	0.077

Model description

Developed by: DataHive AI
Model type: LoRA (PEFT) adapter for automatic speech recognition
Base model: openai/whisper-large-v3
Language: Arabic (dialectal — Hejazi, Najdi, Jordanian, Moroccan Darija)
Task: transcribe (Arabic)
Access terms: gated — granted for evaluation; please don't redistribute the weights (agreed to at request time). No separate license file; the access request is the agreement.

Intended uses & limitations

Direct use. Transcription of conversational, dialectal Arabic in the covered dialects (Hejazi, Najdi, Jordanian, Moroccan Darija), including emotional/expressive speech.

Out-of-scope / use with care.

Not a universally better Arabic model. The advantage is specialization for this domain and these dialects/convention. On neutral, out-of-domain benchmarks (e.g. MSA-adjacent Saudi speech in a third-party convention) frontier systems are competitive, and on some external Moroccan sets a larger open model is ahead. See the whitepaper for the honest, non-cherry-picked comparison.
Dialects and domains outside the training coverage are not guaranteed; results reflect our dialectal convention.
ASR output can contain errors; do not use for high-stakes decisions without human review.

How to get started

The weights are gated — request access on this page first. Once granted:

python
from peft import PeftModel
from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch, torchaudio

BASE = "openai/whisper-large-v3"
ADAPTER = "datahiveai/whisper-large-v3-ar-dialects"

processor = WhisperProcessor.from_pretrained(BASE)
model = WhisperForConditionalGeneration.from_pretrained(BASE, torch_dtype=torch.float16)
model = PeftModel.from_pretrained(model, ADAPTER)   # requires access token with granted access
model = model.merge_and_unload().to("cuda").eval()

wav, sr = torchaudio.load("clip.wav")
if sr != 16000:
    wav = torchaudio.functional.resample(wav, sr, 16000)
feats = processor(wav.squeeze().numpy(), sampling_rate=16000, return_tensors="pt").input_features
feats = feats.to("cuda", torch.float16)

ids = model.generate(feats, language="ar", task="transcribe")
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

Training details

Base: openai/whisper-large-v3 (LoRA / PEFT — only the adapter is trained)
LoRA: r = 64, α = 128, dropout = 0.05, target modules = q_proj, k_proj, v_proj, out_proj
Optimizer / schedule: lr 1e-4, warmup 500 steps, weight decay 0.01
Batch: 8 × grad-accum 2 (effective 16), seed 42, fp16
Data mixture: 40% proprietary corpus, 33.75% SADA, 26.25% MASC (public part held byte-identical between the public-only and public+ours ablation recipes)
Audio augmentation: noise/RIR augmentation applied more heavily to the proprietary corpus than to the already-in-the-wild public data

Evaluation

Test set: proprietary held-out, speaker + script + text-disjoint (~3,354 clips, four dialects) — no speaker, script, or sentence is shared with training, so results reflect genuine generalization.
Metrics: WER and CER, under two normalizers — the Open Universal Arabic ASR Leaderboard normalizer (faithful) and a convention-robust CAMeL-standard normalizer; 95% CIs from 2,000× paired bootstrap.
Honest caveat: part of any model's apparent lead on matched data is orthographic convention; under convention-robust scoring our lead narrows but holds against strong baselines, and is essentially unchanged versus stock (real quality, not spelling). Full tables, external transfer tests, and the losses we don't win are in the whitepaper.

Citation

bibtex
@misc{datahive_whisper_ar_dialects,
  title  = {The Data Moat in Dialectal Arabic Speech Recognition},
  author = {DataHive AI},
  year   = {2026},
  url    = {https://datahive.ai/blog/2026/07/06/the-data-moat-in-dialectal-arabic-speech-recognition/},
  note   = {Whisper-large-v3 + LoRA on a proprietary 4-dialect Arabic corpus}
}

whisper-large-v3-ar-dialects

README

Results (held-out target test · n = 3,354 · speaker/script/text-disjoint)

Model description

Intended uses & limitations

How to get started

Training details

Evaluation

Citation

Data & licensing — get in touch

Explore FriendliAI today

README

Results (held-out target test · n = 3,354 · speaker/script/text-disjoint)

Model description

Intended uses & limitations

How to get started

Training details

Evaluation

Citation

Data & licensing — get in touch