Voho Saudi STT Small
Speech recognition for Saudi Arabic as it is actually spoken: Najdi, Hijazi and Khaleeji dialects, from Voho.
Most Arabic speech models are trained on Modern Standard Arabic: the news, not a phone call from Riyadh. This model is fine-tuned on about 187,000 clips of Saudi dialect speech, and on Saudi speech it makes far fewer mistakes than the model it started from.
Results
Word and character error rate on the Saudi dialect test set (lower is better), before and after fine-tuning:
| Dialect | Clips | WER before | WER after | CER before | CER after |
|---|---|---|---|---|---|
| All Saudi test clips | 4,582 | 103.9% | 39.2% | 69.0% | 16.9% |
| Najdi (Riyadh, central) | 1,704 | 102.8% | 35.7% | 68.7% | 15.3% |
| Hijazi (Jeddah, Makkah) | 809 | 104.3% | 36.3% | 67.9% | 15.4% |
| Khaleeji (Eastern Province, Gulf) | 1,150 | 106.6% | 43.2% | 70.5% | 18.3% |
| Saudi, dialect unlabelled | 762 | 135.8% | 51.2% | 97.3% | 25.1% |
| Modern Standard Arabic | 157 | 52.6% | 33.7% | 29.5% | 12.6% |
Scored after the normalisation standard for Arabic speech recognition: diacritics removed, alef forms unified, ta marbuta and alef maqsura normalised, punctuation removed. Both models were scored identically.
Examples
| Dialect | What was said | What the model wrote |
|---|---|---|
| Khaleeji | الديرة نت الديرة دوت الديرة نت الديرة نت. | هدير هدير عليك برطنا هدير عليك هدير هدير عليك برطنا هدير عليك |
| Najdi | يا ليتك يا فيصل تعرف وش إللي أبغاه. | ليلتك يا فيصل تعرف وش اللي أبغى |
Usage
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="VohoAI/voho-saudi-stt-small")
print(asr("call.wav", generate_kwargs={"language": "arabic", "task": "transcribe"})["text"])
Training
- Base model:
openai/whisper-small(244M parameters, MIT) - Data: SADA, the Saudi Audio Dataset for Arabic published by SDAIA and the National Center for AI, Saudi TV speech with dialect labels. Clips labelled Najdi, Hijazi, Khaleeji, unlabelled Saudi and Modern Standard Arabic were kept. Overlapping multi-speaker clips, non-Saudi dialects and clips over 30 seconds were removed.
- Targets: transcripts with diacritics removed
- Setup: 2 epochs, batch size 32, learning rate 1e-5, bf16, one NVIDIA L4
Licence and intended use
Non-commercial. SADA is licensed CC BY-NC-SA 4.0, so this model is released under the same licence. You may use, share and adapt it for research and non-commercial purposes, with attribution, under the same terms. It is not licensed for commercial use.
For production Saudi Arabic speech recognition and synthesis, including in-Kingdom and on-premise deployment, use the Voho API or the LiveKit plugin.
Limitations
- Trained mostly on television speech. Telephone audio (8 kHz, compressed, noisy) is harder, and accuracy on real calls will be lower than the table above.
- Small model: fast, but less accurate than larger speech models.
- Speaker gender in the training data is skewed male.
- Transcripts are written without diacritics.
Citation
Please cite SADA when using this model:
Saudi Audio Dataset for Arabic (SADA), Saudi Data and AI Authority (SDAIA), 2022.