Downloads · 30 days
9
1% of all-time downloads
zohirjonsharipov/xls-r-uzbek-cv8
xls-r-uzbek-cv8 is a automatic speech recognition model from zohirjonsharipov. Use it when you need speech turned into text. It is set up for transformers. The card lists the license as apache-2.0.
Ushbu model facebook/wav2vec2-xls-r-300m asosida MOZILLA-FOUNDATION/COMMONVOICE80 - UZ datasetidan foydalangan holda Transfer Learning usuli orqali ngramm modeli asosida o'zbek tili uchun fine-tuning qilingan. Model q…
Downloads · 30 days
9
1% of all-time downloads
All-time downloads
618
Public
Parameters
315M
2.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.bin1.3 GB · 45%
From the Hugging Face model README
Ushbu model facebook/wav2vec2-xls-r-300m asosida MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - UZ datasetidan foydalangan holda Transfer Learning usuli orqali ngramm modeli asosida o'zbek tili uchun fine-tuning qilingan. Model quydagi natijalarga erishgan:
Model arxitekturasi haqida ko'prom ma'lumot olish uchun ushbu facebook/wav2vec2-xls-r-300m havola orqali o'ting
Ushbu modelning lugʻati oʻzbek tili zamonaviy lotin alifbosidan iborat boʻlib, tinish belgilari olib tashlangan(https://en.wikipedia.org/wiki/Uzbek_alphabet). Shuni ta'kidlash kerakki, <‘> va <’> belgilar tinish belgisi sifatida hisoblanmaydi, qachonki mana shunday belgilar <o> va <g> dan so'ng kelganda ularni <‘> bilan o‘zgartirilgan.
Dekoder common_voice matniga asoslangan kenlm tili modelidan foydalanadi.
Ushbu model quyidagi foydalanish holatlari uchun foydali bo'lishi kutilmoqda:
Model jonli efirdagi uchrashuvlar yoki ko'rsatuvlarni subtitrini aniqlash uchun kerakli ravishda mos emas va undan Common Voice maʼlumotlar toʻplamiga yoki boshqa hissa qoʻshuvchilarning shaxsiy hayotini xafvga qo'yadigan holatlar uchun ishlatilmasligi kerak.
The 50% of the train common voice official split was used as training data. The 50% of the official dev split was used as validation data, and the full test set was used for final evaluation of the model without LM, while the model with LM was evaluated only on 500 examples from the test set.
The kenlm language model was compiled from the target sentences of the train + other dataset splits.
Training jarayonida quyidagi giperparametrlardan foydalanildi:
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
|---|---|---|---|---|---|
| 3.1401 | 3.25 | 500 | 3.1146 | 1.0 | 1.0 |
| 2.7484 | 6.49 | 1000 | 2.2842 | 1.0065 | 0.7069 |
| 1.0899 | 9.74 | 1500 | 0.5414 | 0.6125 | 0.1351 |
| 0.9465 | 12.99 | 2000 | 0.4566 | 0.5635 | 0.1223 |
| 0.8771 | 16.23 | 2500 | 0.4212 | 0.5366 | 0.1161 |
| 0.8346 | 19.48 | 3000 | 0.3994 | 0.5144 | 0.1102 |
| 0.8127 | 22.73 | 3500 | 0.3819 | 0.4944 | 0.1051 |
| 0.7833 | 25.97 | 4000 | 0.3705 | 0.4798 | 0.1011 |
| 0.7603 | 29.22 | 4500 | 0.3661 | 0.4704 | 0.0992 |
| 0.7424 | 32.47 | 5000 | 0.3529 | 0.4577 | 0.0957 |
| 0.7251 | 35.71 | 5500 | 0.3410 | 0.4473 | 0.0928 |
| 0.7106 | 38.96 | 6000 | 0.3401 | 0.4428 | 0.0919 |
| 0.7027 | 42.21 | 6500 | 0.3355 | 0.4353 | 0.0905 |
| 0.6927 | 45.45 | 7000 | 0.3308 | 0.4296 | 0.0885 |
| 0.6828 | 48.7 | 7500 | 0.3246 | 0.4204 | 0.0863 |
| 0.6706 | 51.95 | 8000 | 0.3250 | 0.4233 | 0.0868 |
| 0.6629 | 55.19 | 8500 | 0.3264 | 0.4159 | 0.0849 |
| 0.6556 | 58.44 | 9000 | 0.3213 | 0.4100 | 0.0835 |
| 0.6484 | 61.69 | 9500 | 0.3182 | 0.4124 | 0.0837 |
| 0.6407 | 64.93 | 10000 | 0.3171 | 0.4050 | 0.0825 |
| 0.6375 | 68.18 | 10500 | 0.3150 | 0.4039 | 0.0822 |
| 0.6363 | 71.43 | 11000 | 0.3129 | 0.3991 | 0.0810 |
| 0.6307 | 74.67 | 11500 | 0.3114 | 0.3986 | 0.0807 |
| 0.6232 | 77.92 | 12000 | 0.3103 | 0.3895 | 0.0790 |
| 0.6216 | 81.17 | 12500 | 0.3086 | 0.3891 | 0.0790 |
| 0.6174 | 84.41 | 13000 | 0.3082 | 0.3881 | 0.0785 |
| 0.6196 | 87.66 | 13500 | 0.3059 | 0.3875 | 0.0782 |
| 0.6174 | 90.91 | 14000 | 0.3084 | 0.3862 | 0.0780 |
| 0.6169 | 94.16 | 14500 | 0.3070 | 0.3860 | 0.0779 |
| 0.6166 | 97.4 | 15000 | 0.3066 | 0.3855 | 0.0778 |