Update README.md
Browse files
README.md
CHANGED
|
@@ -4,8 +4,37 @@ language:
|
|
| 4 |
metrics:
|
| 5 |
- wer
|
| 6 |
- bleu
|
|
|
|
| 7 |
tags:
|
| 8 |
- ASR
|
| 9 |
- Error Correction
|
| 10 |
- Crossmodal
|
| 11 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
metrics:
|
| 5 |
- wer
|
| 6 |
- bleu
|
| 7 |
+
- google_bleu
|
| 8 |
tags:
|
| 9 |
- ASR
|
| 10 |
- Error Correction
|
| 11 |
- Crossmodal
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
### Model Description
|
| 15 |
+
|
| 16 |
+
Pre-Training Settings:
|
| 17 |
+
|
| 18 |
+
166k samples from Common Voice 13.0 was recognized by Whisper tiny.en.
|
| 19 |
+
|
| 20 |
+
1,000 random samples was selected as the test set, and the rest for training and validation with an 80%-20% split
|
| 21 |
+
|
| 22 |
+
- Batch size: 256
|
| 23 |
+
|
| 24 |
+
- Initial learning rate: 1e-5
|
| 25 |
+
|
| 26 |
+
- Adam optimizer
|
| 27 |
+
|
| 28 |
+
- 30 epochs
|
| 29 |
+
|
| 30 |
+
- Cross-entropy loss
|
| 31 |
+
|
| 32 |
+
- Best checkpoint saved based on WER as the evaluation metric
|
| 33 |
+
|
| 34 |
+
- Decoding is performed using beam search with a size of 5
|
| 35 |
+
|
| 36 |
+
- S2S backbone model adopted from ''[Exploring data augmentation for code generation tasks](https://aclanthology.org/2023.findings-eacl.114/)''.
|
| 37 |
+
|
| 38 |
+
Continue-Training Setting:
|
| 39 |
+
|
| 40 |
+
- 2 epochs for gold-gold to prevent the over-correction problem on ''[Ted talk data](https://cris.fbk.eu/bitstream/11582/104409/1/WIT3-EAMT2012.pdf)''
|