oddadmix

whisper-large-v3-arabic-dialectal-v2

README

License: apache-2.0

Results (balanced test split ~2,600 clips, `clean_text` WER/CER)

Table with columns: WER, CER
	WER	CER
This model (v2)	0.3199	0.1203

Per-dialect accuracy (v3 balanced test, `clean_text`)

WER / CER on each dialect's 200 held-out clips, easiest to hardest. Gulf/Levantine transcribe best; Maghrebi (Tunisian/Algerian/Moroccan) is hardest — consistent across every model.

Table with columns: Dialect, WER, CER, n
Dialect	WER	CER	n
Saudi/Gulf (`sa`)	0.169	0.048	200
Iraqi (`iq`)	0.254	0.068	200
Egyptian (`eg`)	0.271	0.134	200
Syrian (`sy`)	0.272	0.096	200
Palestinian (`ps`)	0.301	0.137	200
Yemeni (`ye`)	0.304	0.112	200
Lebanese (`lb`)	0.317	0.113	200
Moroccan (`ma`)	0.340	0.099	200
Libyan (`ly`)	0.343	0.132	200
Bahraini (`bh`)	0.348	0.136	200
Algerian (`dz`)	0.386	0.156	200
Sudanese (`sd`)	0.399	0.160	200
Tunisian (`tn`)	0.478	0.193	200
Overall	0.320	0.120	2600

v2 comparison — all models on the balanced dataset

Same dialect-balanced test split (~2,600 clips, oddadmix/lahgtna-v3-small), same clean_text scoring, seed 42.

Table with columns: Model, Params, WER, CER
Model	Params	WER	CER
whisper-large-v3-v2	1.54B	0.320	0.120
whisper-large-v3-turbo-v2	809M	0.332	0.125
whisper-medium-v2	769M	0.342	0.128
whisper-small-v2	244M	0.403	0.153
nemotron-3.5-asr-v2 (streaming)

whisper-large-v3-turbo-v2 is the best; Whisper fine-tunes furthest. nemotron is the streaming option. (v1 models were trained on the earlier augmentation-expanded set.)

Examples — base vs. fine-tuned (one clip per dialect)

13 held-out test clips, one per dialect. Base is openai/whisper-large-v3 untouched; fine-tuned (v2) is this model. Text is clean_text-normalized (undiacritized), matching the WER scoring. Press play to listen.

Table with columns: Dialect, Audio, Reference, Base, Fine-tuned
Dialect	Reference	Base	Fine-tuned
Egyptian (`eg`)	لا، في كافيه هنا كده جنب الدار، مع أحمد خطيبها، وزمانها جاي.	لافك فيه هنا كده جنب الدور مع أحمد خطبها وزمنها دي	لافكة فيه هنا كده جنب الدار، مع أحمد خطيبها وزمنها جاي.
Lebanese (`lb`)	ما كون مأمنة حالي. ما أقدر اتحكم. امم. في عندي خوف، إيه.	ما أكون مأمنة حالي ما أقدر أتحكم في عندي خوف	ما أكون مأمنة حالي، ما أقدر أتحكم. إم. في عندي خوف، إيه.
Syrian ()

Full fine-tuning code, dependencies, and lessons are in the v1 repo (train.py, normalize.py, FINETUNE.md). This v2 uses the same recipe on the balanced dataset.

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

oddadmix

Model Tree

Base

openai/whisper-large-v3

Fine-tuned

this model

Input Modalities

Audio

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Results (balanced test split ~2,600 clips, `clean_text` WER/CER)

Table with columns: WER, CER
	WER	CER
This model (v2)	0.3199	0.1203

Per-dialect accuracy (v3 balanced test, `clean_text`)

WER / CER on each dialect's 200 held-out clips, easiest to hardest. Gulf/Levantine transcribe best; Maghrebi (Tunisian/Algerian/Moroccan) is hardest — consistent across every model.

Table with columns: Dialect, WER, CER, n
Dialect	WER	CER	n
Saudi/Gulf (`sa`)	0.169	0.048	200
Iraqi (`iq`)	0.254	0.068	200
Egyptian (`eg`)	0.271	0.134	200
Syrian (`sy`)	0.272	0.096	200
Palestinian (`ps`)	0.301	0.137	200
Yemeni (`ye`)	0.304	0.112	200
Lebanese (`lb`)	0.317	0.113	200
Moroccan (`ma`)	0.340	0.099	200
Libyan (`ly`)	0.343	0.132	200
Bahraini (`bh`)	0.348	0.136	200
Algerian (`dz`)	0.386	0.156	200
Sudanese (`sd`)	0.399	0.160	200
Tunisian (`tn`)	0.478	0.193	200
Overall	0.320	0.120	2600

v2 comparison — all models on the balanced dataset

Same dialect-balanced test split (~2,600 clips, oddadmix/lahgtna-v3-small), same clean_text scoring, seed 42.

Table with columns: Model, Params, WER, CER
Model	Params	WER	CER
whisper-large-v3-v2	1.54B	0.320	0.120
whisper-large-v3-turbo-v2	809M	0.332	0.125
whisper-medium-v2	769M	0.342	0.128
whisper-small-v2	244M	0.403	0.153
nemotron-3.5-asr-v2 (streaming)

whisper-large-v3-turbo-v2 is the best; Whisper fine-tunes furthest. nemotron is the streaming option. (v1 models were trained on the earlier augmentation-expanded set.)

Examples — base vs. fine-tuned (one clip per dialect)

Table with columns: Dialect, Audio, Reference, Base, Fine-tuned
Dialect	Reference	Base	Fine-tuned
Egyptian (`eg`)	لا، في كافيه هنا كده جنب الدار، مع أحمد خطيبها، وزمانها جاي.	لافك فيه هنا كده جنب الدور مع أحمد خطبها وزمنها دي	لافكة فيه هنا كده جنب الدار، مع أحمد خطيبها وزمنها جاي.
Lebanese (`lb`)	ما كون مأمنة حالي. ما أقدر اتحكم. امم. في عندي خوف، إيه.	ما أكون مأمنة حالي ما أقدر أتحكم في عندي خوف	ما أكون مأمنة حالي، ما أقدر أتحكم. إم. في عندي خوف، إيه.
Syrian ()

Full fine-tuning code, dependencies, and lessons are in the v1 repo (train.py, normalize.py, FINETUNE.md). This v2 uses the same recipe on the balanced dataset.

whisper-large-v3-arabic-dialectal-v2

README

Results (balanced test split ~2,600 clips, clean_text WER/CER)

Per-dialect accuracy (v3 balanced test, clean_text)

v2 comparison — all models on the balanced dataset

Examples — base vs. fine-tuned (one clip per dialect)

Explore FriendliAI today

README

Results (balanced test split ~2,600 clips, clean_text WER/CER)

Per-dialect accuracy (v3 balanced test, clean_text)

v2 comparison — all models on the balanced dataset

Examples — base vs. fine-tuned (one clip per dialect)

Results (balanced test split ~2,600 clips, `clean_text` WER/CER)

Per-dialect accuracy (v3 balanced test, `clean_text`)

Results (balanced test split ~2,600 clips, `clean_text` WER/CER)

Per-dialect accuracy (v3 balanced test, `clean_text`)