videodanchik's picture
Update README.md
2e88d7e
|
raw
history blame
898 Bytes
---
datasets:
- Tedlium3
language:
- en
license: cc-by-nc-nd-3.0
metrics:
-
name: "dev"
type: wer
value: 5.60
-
name: "test"
type: wer
value: 5.33
tags:
- k2
- icefall
- automatic-speech-recognition
---
# Tedlium3
This is an ASR ctc attention encoder-decoder recipe for the Tedlium3 corpus.
The encoder is the re-worked conformer model.
Language models are downloaded from https://kaldi-asr.org/models/m5
## Performance Record
### conformer_ctc2
| Decodeing method | dev WER | test WER |
|--------------------------|------------|---------|
| ctc-decoding | 6.45 | 5.96 |
| 1best | 5.92 | 5.51 |
| whole-lattice-rescoring | 5.96 | 5.47 |
| attention-decoder | 5.60 | 5.33 |
See the [recipe](https://github.com/k2-fsa/icefall/tree/master/egs/tedlium3/ASR/conformer_ctc2) for details.