Update README.md

README.md

@@ -8,41 +8,60 @@ tags:
 - generated_from_trainer
 - audio
 - speech
+- belarusian
 datasets:
 - sarulab-speech/commonvoice22_sidon
+metrics:
+- wer
 model-index:
 - name: Whisper Small Belarusian Custom
-  results: []
+  results:
+  - task:
+      type: automatic-speech-recognition
+      name: Speech Recognition
+    dataset:
+      name: Common Voice 22 Sidon (Belarusian)
+      type: sarulab-speech/commonvoice22_sidon
+      config: be
+      split: test
+    metrics:
+    - type: wer
+      value: 27.27
+      name: Word Error Rate
 ---
 
 # Whisper Small Belarusian (Common Voice 22 Sidon)
 
-## 🇺🇸 English
-This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) on the Belarusian dataset from **Common Voice 22 Sidon** (`sarulab-speech/commonvoice22_sidon`). It achieves improved performance on Belarusian speech recognition compared to the base model.
+A fine-tuned version of openai/whisper-small optimized for Belarusian speech recognition. This model significantly outperforms both the base Whisper Small and even the much larger Whisper Large V3 on Belarusian speech.
 
-**Metrics (approximate):**
-- **WER (Word Error Rate):** ~20% (after 1200 steps)
-- **Loss:** ~0.21
+---
 
-### Usage
-```python
-from transformers import pipeline
+## Benchmark Results
 
-pipe = pipeline("automatic-speech-recognition", model="aleton/whisper-small-be-custom")
-result = pipe("audio_file.mp3")
-print(result["text"])
-```
+| Model | Parameters | WER (lower is better) |
+|:------|:----------:|----------------------:|
+| openai/whisper-small (base) | 244M | 92.21% |
+| openai/whisper-large-v3 | 1550M | 63.64% |
+| **This model** | **244M** | **27.27%** |
+
+**Key finding:** This fine-tuned Small model outperforms Whisper Large V3 by 36.37 percentage points, while being 6x smaller in size.
 
 ---
 
-## 🇷🇺 Русский
-Эта модель является дообученной (fine-tuned) версией [openai/whisper-small](https://huggingface.co/openai/whisper-small) на наборе данных **Common Voice 22 Sidon** (белорусский язык). Модель показывает лучшие результаты распознавания белорусской речи по сравнению с базовой версией.
+## Training Details
+
+- **Base Model:** openai/whisper-small
+- **Dataset:** sarulab-speech/commonvoice22_sidon (Belarusian subset)
+- **Training Steps:** 1700
+- **Learning Rate:** 1e-5
+- **Batch Size:** 8
+- **Final Loss:** ~0.21
+- **Framework:** PyTorch, Transformers, Accelerate
+
+---
 
-**Метрики (примерные):**
-- **WER (Процент ошибок):** ~20% (после 1200 шагов обучения)
-- **Loss (Ошибка):** ~0.21
+## Usage
 
-### Как использовать
 ```python
 from transformers import pipeline
 
@@ -51,31 +70,49 @@ result = pipe("audio_file.mp3")
 print(result["text"])
 ```
 
+For longer audio files:
+
+```python
+result = pipe("long_audio.mp3", chunk_length_s=30, batch_size=8)
+print(result["text"])
+```
+
 ---
 
-## 🇧🇾 Беларуская
-Гэтая мадэль з'яўляецца данавучанай (fine-tuned) версіяй [openai/whisper-small](https://huggingface.co/openai/whisper-small) на наборы дадзеных **Common Voice 22 Sidon** (беларуская мова). Мадэль паказвае лепшыя вынікі распазнавання беларускай мовы ў параўнанні з базавай версіяй.
+## Description
 
-**Метрыкі (прыкладныя):**
-- **WER (Працэнт памылак):** ~20% (пасля 1200 крокаў навучання)
-- **Loss (Страты):** ~0.21
+### English
 
-### Як выкарыстоўваць
-```python
-from transformers import pipeline
+This model demonstrates that targeted fine-tuning on language-specific data can dramatically improve performance for low-resource languages. The base Whisper models struggle with Belarusian due to limited representation in the original training data. Through fine-tuning on Common Voice 22 Sidon, this model achieves a 64.94 percentage point improvement over the base Small model and a 36.37 percentage point improvement over the Large V3 model.
 
-pipe = pipeline("automatic-speech-recognition", model="aleton/whisper-small-be-custom")
-result = pipe("audio_file.mp3")
-print(result["text"])
-```
+### Русский
+
+Эта модель демонстрирует, что целенаправленное дообучение на языковых данных может значительно улучшить качество распознавания для малоресурсных языков. Базовые модели Whisper плохо справляются с белорусским языком из-за ограниченного представления в исходных обучающих данных. Благодаря дообучению на Common Voice 22 Sidon, эта модель показывает улучшение на 64.94 п.п. по сравнению с базовой Small моделью и на 36.37 п.п. по сравнению с Large V3.
+
+### Беларуская
+
+Гэтая мадэль дэманструе, што мэтанакіраванае данавучанне на моўных дадзеных можа значна палепшыць якасць распазнавання для маларэсурсных моў. Базавыя мадэлі Whisper дрэнна спраўляюцца з беларускай мовай з-за абмежаванага прадстаўніцтва ў зыходных навучальных дадзеных. Дзякуючы данавучанню на Common Voice 22 Sidon, гэтая мадэль паказвае паляпшэнне на 64.94 п.п. у параўнанні з базавай Small мадэллю і на 36.37 п.п. у параўнанні з Large V3.
 
 ---
 
-### Training parameters / Параметры обучения / Параметры навучання
-- **Base Model:** openai/whisper-small
-- **Dataset:** sarulab-speech/commonvoice22_sidon (be)
-- **Learning Rate:** 1e-5
-- **Batch Size:** 8
-- **Steps:** 1200
-- **Framework:** PyTorch, Transformers, Accelerate
-```
+## Limitations
+
+- Optimized specifically for Belarusian; performance on other languages may be degraded compared to the base model
+- Trained on Common Voice data, which may not fully represent all dialects or acoustic conditions
+- Best results on clear audio with minimal background noise
+
+---
+
+## Citation
+
+If you use this model, please cite:
+
+```bibtex
+@misc{whisper-small-be-custom,
+  author = {aleton},
+  title = {Whisper Small Belarusian Custom},
+  year = {2024},
+  publisher = {Hugging Face},
+  url = {https://huggingface.co/aleton/whisper-small-be-custom}
+}
+```