valiyevfagan/whisper-small-az

Whisper Small Azerbaijani ASR

Fine-tuned Whisper Small model for Azerbaijani automatic speech recognition.

Model Description

This model is a fine-tuned version of openai/whisper-small on Azerbaijani speech data. It converts spoken Azerbaijani audio into text with high accuracy.

Model Details

• Developed by: Fagan Valiyev
• Model type: Automatic Speech Recognition (ASR)
• Language: Azerbaijani (az)
• Base model: openai/whisper-small
• Parameters: ~244M
• License: Apache 2.0 (following Whisper's license)

Out-of-Scope Use

The model should not be used for:

• Real-time critical applications without human review
• Legal or medical transcription without verification
• Surveillance or privacy-violating applications
• Languages other than Azerbaijani

Training Data

The model was fine-tuned on Azerbaijani speech datasets. Training details:

• Training checkpoint: 160
• Traing data: common_voice_17_0
• Audio preprocessing: 16kHz sampling rate
• Data augmentation: Applied standard audio augmentation techniques

Training Results

The model was trained for 1600 steps with the following progression:

Step	Training Loss	Validation Loss	WER (%)
20	1.152000	0.774133	45.41
40	0.745000	0.482266	34.78
60	0.321600	0.326407	24.64
80	0.218100	0.209071	12.56
100	0.068300	0.083807	6.76
120	0.077400	0.046254	4.35
140	0.021100	0.021606	0.97
160	0.015900	0.017379	0.48

Final Performance (Checkpoint 160):

• Word Error Rate (WER): 0.48%
• Validation Loss: 0.017379
• Training Loss: 0.015900

Evaluation

The model achieves excellent performance on Azerbaijani speech recognition with a WER of less than 1%.

Best results are achieved with:

• Clear audio with minimal background noise
• Single speaker recordings
• Standard Azerbaijani dialect
• Audio sampled at 16kHz

Limitations and Bias

Limitations

• Performance degrades with heavy background noise
• May struggle with strong regional dialects
• Less accurate on very short utterances (<1 second)
• Mixed-language speech may not transcribe accurately
• Technical or domain-specific terminology may have lower accuracy

Bias

• The model may perform differently across:
  • Gender (male vs. female voices)
  • Age groups (children, adults, elderly)
  • Regional dialects and accents
  • Speaking styles (formal vs. informal)

Users should validate outputs, especially for critical applications

Model Card Authors

Fagan Valiyev

---

Additional Resources:

• 🎙️ Try the Streamlit Demo
• 📖 Whisper Documentation
• 🤗 Transformers Library