Upload README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
datasets:
|
| 3 |
+
- Bingsu/zeroth-korean
|
| 4 |
+
language:
|
| 5 |
+
- ko
|
| 6 |
+
metrics:
|
| 7 |
+
- cer
|
| 8 |
+
- wer
|
| 9 |
+
base_model:
|
| 10 |
+
- openai/whisper-large-v3-turbo
|
| 11 |
+
pipeline_tag: automatic-speech-recognition
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
## Description
|
| 15 |
+
Fine-tuning Whisper Large V3 Turbo on zeroth Korean dataset.
|
| 16 |
+
|
| 17 |
+
## Dataset split:
|
| 18 |
+
- The test dataset from Korean zeroth is divided to test and validation -> 50% validation, 50% test
|
| 19 |
+
- Train set duration: 206 hours 43 minutes
|
| 20 |
+
- Validation set duration: 2 hours 22 minutes
|
| 21 |
+
- Test set duration: 2 hours 22 minutes
|
| 22 |
+
|
| 23 |
+
## Results:
|
| 24 |
+
- initial validation WER: 26.26%
|
| 25 |
+
- final validation WER: 4.90%
|
| 26 |
+
- initial validation CER: 6.67%
|
| 27 |
+
- final validation CER: 1.78%
|
| 28 |
+
- initial test WER: 26.75%
|
| 29 |
+
- final test WER: 4.89%
|
| 30 |
+
- initial test CER: 7.58%
|
| 31 |
+
- final test CER: 2.06%
|
| 32 |
+
|
| 33 |
+
## Notes
|
| 34 |
+
- Models did not converge, better results are possible.
|