Update README.md
Browse files
README.md
CHANGED
|
@@ -27,3 +27,37 @@ configs:
|
|
| 27 |
- split: train
|
| 28 |
path: data/train-*
|
| 29 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
- split: train
|
| 28 |
path: data/train-*
|
| 29 |
---
|
| 30 |
+
|
| 31 |
+
# Tibetan ASR Model Comparison: Garchen Rinpoche Dataset
|
| 32 |
+
|
| 33 |
+
## Dataset Information
|
| 34 |
+
|
| 35 |
+
- Content: Garchen Rinpoche Tibetan teachings
|
| 36 |
+
- Samples: 552 audio segments
|
| 37 |
+
- Duration: 1.21 hours total
|
| 38 |
+
- Metrics used: Character Error Rate (CER), Word Error Rate (WER), Syllable Error Rate (SER)
|
| 39 |
+
|
| 40 |
+
## Model Performance Summary
|
| 41 |
+
|
| 42 |
+
### General STT Model
|
| 43 |
+
- Model: [OpenPecha General Tibetan STT](https://huggingface.co/openai/whisper-large-v3)
|
| 44 |
+
- Character Error Rate: 27.53%
|
| 45 |
+
- Word Error Rate: 57.24%
|
| 46 |
+
- Syllable Error Rate: 51.15%
|
| 47 |
+
- Major Error Type: Word substitutions (75%)
|
| 48 |
+
|
| 49 |
+
### Situ Rinpoche Model
|
| 50 |
+
- Model: [OpenPecha Situ Rinpoche ASR](https://huggingface.co/openpecha/Situ_rinpoche_asr)
|
| 51 |
+
- Character Error Rate: 28.92%
|
| 52 |
+
- Word Error Rate: 60.68%
|
| 53 |
+
- Syllable Error Rate: 53.68%
|
| 54 |
+
- Major Error Type: Word substitutions (75%)
|
| 55 |
+
|
| 56 |
+
### Dilgo Khyentse Model
|
| 57 |
+
- Model: [OpenPecha Dilgo Khyentse ASR](https://huggingface.co/ganga4364/kr_wav2vec2_v2.137000)
|
| 58 |
+
- Character Error Rate: 65.96%
|
| 59 |
+
- Word Error Rate: 92.63%
|
| 60 |
+
- Syllable Error Rate: 91.55%
|
| 61 |
+
- Major Error Type: High proportion of deletions (33%)
|
| 62 |
+
|
| 63 |
+
### Detailed Documentation [here](https://forum.openpecha.org/t/stt-evaluation-report-on-garchen-rinpoche-by-different-models/323)
|