Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model: openai/whisper-large-v3
|
| 3 |
+
datasets:
|
| 4 |
+
- bn
|
| 5 |
+
language: bn
|
| 6 |
+
library_name: transformers
|
| 7 |
+
license: apache-2.0
|
| 8 |
+
model-index:
|
| 9 |
+
- name: Finetuned openai/whisper-large-v3 on Bengali
|
| 10 |
+
results:
|
| 11 |
+
- task:
|
| 12 |
+
type: automatic-speech-recognition
|
| 13 |
+
name: Speech-to-Text
|
| 14 |
+
dataset:
|
| 15 |
+
name: Common Voice (Bengali)
|
| 16 |
+
type: common_voice
|
| 17 |
+
metrics:
|
| 18 |
+
- type: wer
|
| 19 |
+
value: 9.651
|
| 20 |
+
---
|
| 21 |
+
|
| 22 |
+
# Finetuned openai/whisper-large-v3 on 21409 Bengali training audio samples from cv-corpus-21.0-2025-03-14/bn.
|
| 23 |
+
|
| 24 |
+
This model was created from the Mozilla.ai Blueprint:
|
| 25 |
+
[speech-to-text-finetune](https://github.com/mozilla-ai/speech-to-text-finetune).
|
| 26 |
+
|
| 27 |
+
## Evaluation results on 9363 audio samples of Bengali:
|
| 28 |
+
|
| 29 |
+
### Baseline model (before finetuning) on Bengali
|
| 30 |
+
- Word Error Rate (Normalized): 55.463
|
| 31 |
+
- Word Error Rate (Orthographic): 83.344
|
| 32 |
+
- Character Error Rate (Normalized): 35.66
|
| 33 |
+
- Character Error Rate (Orthographic): 40.754
|
| 34 |
+
- Loss: 0.567
|
| 35 |
+
|
| 36 |
+
### Finetuned model (after finetuning) on Bengali
|
| 37 |
+
- Word Error Rate (Normalized): 9.651
|
| 38 |
+
- Word Error Rate (Orthographic): 24.288
|
| 39 |
+
- Character Error Rate (Normalized): 4.876
|
| 40 |
+
- Character Error Rate (Orthographic): 6.312
|
| 41 |
+
- Loss: 0.092
|