Chillarmo/whisper-base-armenian

Whisper-Base Fine-tuned for Armenian ASR

This model is a fine-tuned version of OpenAI's Whisper-base on the Common Voice 20 Armenian dataset for automatic speech recognition.

Training Results

The model was trained for 5.34 epochs with the following final results:

Metric	Value
Training Loss	0.122
Training Runtime	10,924 seconds (≈3.03 hours)
Training Samples/Second	7.32
Training Steps/Second	0.46
Total Training Steps	5,000
Epochs	5.34

Evaluation Results

Metric	Value
Evaluation Loss	0.201
Word Error Rate (WER)	33.19%
Character Error Rate (CER)	6.98%
BLEU Score	47.71
Exact Match	16.50%
Average Prediction Length	7.69 tokens
Average Label Length	7.77 tokens
Length Ratio	0.989
Evaluation Runtime	1,590 seconds (≈26.5 minutes)
Evaluation Samples/Second	3.68
Evaluation Steps/Second	0.46

Model Details

• Base Model: openai/whisper-base
• Language: Armenian (hy)
• Dataset: Chillarmo/common_voice_20_armenian
• License: Apache 2.0

Notes

During model loading, there were missing keys in the checkpoint: ['proj_out.weight']. This is a common occurrence when fine-tuning Whisper models and typically doesn't affect performance significantly.