Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,49 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-4.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
language:
|
| 4 |
+
- cs
|
| 5 |
+
metrics:
|
| 6 |
+
- cer
|
| 7 |
+
- wer
|
| 8 |
+
---
|
| 9 |
+
# Small fastconformer CTC ASR model
|
| 10 |
+
|
| 11 |
+
A small fastconformer CTC ASR model for light, especially low to medium vocabulary applications.
|
| 12 |
+
|
| 13 |
+
## Model Details
|
| 14 |
+
|
| 15 |
+
### Model Description
|
| 16 |
+
|
| 17 |
+
- **Developed by:** Vojtěch Drábek
|
| 18 |
+
- **Model type:** Fastconformer CTC
|
| 19 |
+
- **Language(s) (NLP):** Czech
|
| 20 |
+
- **License:** CC-BY-4.0
|
| 21 |
+
|
| 22 |
+
## Evaluation
|
| 23 |
+
|
| 24 |
+
### Testing Data
|
| 25 |
+
|
| 26 |
+
Mozilla Common Voice Czech 21.0 Test Set
|
| 27 |
+
|
| 28 |
+
### Results
|
| 29 |
+
|
| 30 |
+
#### Greedy decoding
|
| 31 |
+
|
| 32 |
+
##### Full model
|
| 33 |
+
WER: 15.0
|
| 34 |
+
CER: 5.2
|
| 35 |
+
|
| 36 |
+
##### Quantized model
|
| 37 |
+
WER: 15.5
|
| 38 |
+
CER: 5.4
|
| 39 |
+
|
| 40 |
+
#### Been search decoding
|
| 41 |
+
|
| 42 |
+
##### General language model
|
| 43 |
+
|
| 44 |
+
- pyctcdecode
|
| 45 |
+
- beam width 16
|
| 46 |
+
- 5-gram, 500k words, 992 MB
|
| 47 |
+
|
| 48 |
+
WER: 11.6
|
| 49 |
+
CER: 5.8
|