Downloads · 30 days
28
19% of all-time downloads
MohamedGomaa30/stt_ar_fastconformer_casablanca_nadi
stt_ar_fastconformer_casablanca_nadi is a automatic speech recognition model from MohamedGomaa30. Use it when you need speech turned into text. It is set up for nemo. The card lists the license as cc-by-4.0.
This model is a fine-tuned version of nvidia/sttarfastconformerhybridlargepcdv1.0 (114.6M parameters) on the Casablanca-NADI dialectal Arabic dataset.
Downloads · 30 days
28
19% of all-time downloads
All-time downloads
147
Public
Repo size
459 MB
Likes
1
Public
Click a slice to open those files.
.nemo459 MB · 100%
From the Hugging Face model README
This model is a fine-tuned version of nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0 (114.6M parameters) on the Casablanca-NADI dialectal Arabic dataset.
Casablanca-NADI – 8 Arabic dialects from the NADI shared task:
| Dialect | Train | Validation |
|---|---|---|
| Algeria | 1,482 | 110 |
| Egypt | 1,618 | 106 |
| Jordan | 1,960 | 148 |
| Mauritania | 1,528 | 114 |
| Morocco | 2,094 | 159 |
| Palestine | 1,294 | 95 |
| UAE | 1,534 | 116 |
| Yemen | 1,172 | 107 |
| Total | 12,682 | 955 |
Audio resampled to 16 kHz mono, max duration 20 seconds.
| Metric | Before Fine-tuning | After Fine-tuning | Improvement |
|---|---|---|---|
| WER | 68.58% | 60.54% | -8.04% |
| CER | 30.28% | 24.92% | -5.36% |
| Dialect | Samples | WER Before (base model) | WER After | CER Before (base model) | CER After |
|---|---|---|---|---|---|
| Algeria | 110 | 79.56% | 71.74% | 40.72% | 31.94% |
| Egypt | 106 | 58.14% | 47.79% | 25.13% | 22.45% |
| Jordan | 148 | 46.42% | 38.62% | 14.25% | 12.94% |
| Mauritania | 114 | 81.53% | 77.62% | 43.42% | 36.35% |
| Morocco | 159 | 87.11% | 75.08% | 44.17% | 31.38% |
| Palestine | 95 | 58.06% | 45.50% | 18.26% | 15.42% |
| UAE | 116 | 65.38% | 55.14% | 24.86% | 20.79% |
| Yemen | 107 | 67.82% | 69.72% | 27.13% | 25.93% |
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(
"MohamedGomaa30/stt_ar_fastconformer_casablanca_nadi"
)
# Transcribe a single audio file
transcription = model.transcribe(["path/to/audio.wav"])
print(transcription)
import torch
from nemo.collections.asr.models import EncDecHybridRNNTCTCBPEModel
device = "cuda" if torch.cuda.is_available() else "cpu"
model = EncDecHybridRNNTCTCBPEModel.from_pretrained(
"MohamedGomaa30/stt_ar_fastconformer_casablanca_nadi"
)
model = model.to(device)
model.eval()
stt_ar_fastconformer_hybrid_large_pcd_v1.0