File size: 2,238 Bytes
24a9843 13f4341 d6a4b3d 24a9843 dd0371d 0e52cef 24a9843 68acc64 24a9843 0e52cef |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 |
---
license: apache-2.0
language:
- ko
library_name: nemo
pipeline_tag: automatic-speech-recognition
tags:
- conformer-ctc
metrics:
- wer
---
# Conformer-ctc-medium-ko
ν΄λΉ λͺ¨λΈμ [RIVA Conformer ASR Korean](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/tao/models/speechtotext_ko_kr_conformer)μ AI hub datasetμ λν΄ νμΈνλμ μ§ννμ΅λλ€. <br>
Conformer κΈ°λ°μ λͺ¨λΈμ whisperμ κ°μ attention κΈ°λ° λͺ¨λΈκ³Ό λ¬λ¦¬ streamingμ μ§ννμ¬λ μ±λ₯μ΄ ν¬κ² λ¨μ΄μ§μ§ μκ³ , μλκ° λΉ λ₯΄λ€λ μ₯μ μ΄ μμ΅λλ€.<br>
V100 GPUμμλ RTFκ° 0.05, CPU(7 cores)μμλ 0.35 μ λ λμ€λ κ²μ νμΈν μ μμμ΅λλ€.<br>
μ€λμ€ chunk size 2μ΄μ streaming ν
μ€νΈμμλ μ 체 μ€λμ€λ₯Ό λ£λ κ²μ λΉν΄μλ 20% μ λ μ±λ₯μ νκ° μμΌλ μΆ©λΆν μ¬μ©ν μ μλ μ±λ₯μ
λλ€.<br>
μΆκ°λ‘ open domainμ΄ μλ κ³ κ° μλ μμ±κ³Ό κ°μ domainμμλ kenlmμ μΆκ°νμμ λ WER 13.45μμ WER 5.27λ‘ ν¬κ² μ±λ₯ ν₯μμ΄ μμμ΅λλ€.<br>
νμ§λ§ κ·Έ μΈμ domainμμλ kenlmμ μΆκ°κ° ν° μ±λ₯ ν₯μμΌλ‘ μ΄μ΄μ§μ§ μμμ΅λλ€.
Streaming μ½λμ Denoise modelμ΄ ν¬ν¨λ μ½λλ μλ κΉνμμ νμΈν μ μμ΅λλ€.
[https://github.com/SUNGBEOMCHOI/Korean-Streaming-ASR](https://github.com/SUNGBEOMCHOI/Korean-Streaming-ASR)
### Training results
| Training Loss | Epoch | Wer |
|:-------------:|:-----:|:-------:|
| 9.09 | 1.0 | 11.51 |
### dataset
| λ°μ΄ν°μ
μ΄λ¦ | λ°μ΄ν° μν μ(train/test) |
| --- | --- |
| κ³ κ°μλμμ± | 2067668/21092 |
| νκ΅μ΄ μμ± | 620000/3000 |
| νκ΅μΈ λν μμ± | 2483570/142399 |
| μμ λνμμ±(μΌλ°λ¨λ
) | 1886882/263371 |
| λ³΅μ§ λΆμΌ μ½μΌν° μλ΄λ°μ΄ν° | 1096704/206470 |
| μ°¨λλ΄ λν λ°μ΄ν° | 2624132/332787 |
| λͺ
λ Ήμ΄ μμ±(λ
ΈμΈλ¨μ¬) | 137467/237469 |
| μ 체 | 10916423(13946μκ°)/1206588(1474μκ°) |
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 16
- eval_batch_size: 16
- num_train_epoch: 1
- sample_rate: 16000
- max_duration: 20.0 |