luomingshuang
commited on
Commit
•
7dffd08
1
Parent(s):
6eee019
Update README.md
Browse files
README.md
CHANGED
@@ -1 +1,34 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Pre-trained TDNN-LiGRU-CTC models for the TIMIT dataset with icefall.
|
2 |
+
The model was trained on full [TIMIT](https://data.deepai.org/timit.zip) with the scripts in [icefall](https://github.com/k2-fsa/icefall).
|
3 |
+
See (https://github.com/k2-fsa/icefall/tree/master/egs/timit/ASR/tdnn_ligru_ctc) for more details of this model.
|
4 |
+
## How to use
|
5 |
+
See (https://github.com/k2-fsa/icefall/blob/master/egs/timit/ASR/tdnn_ligru_ctc/Pre-trained.md)
|
6 |
+
## Training procedure
|
7 |
+
The main repositories are list below, we will update the training and decoding scripts with the update of version.
|
8 |
+
k2: https://github.com/k2-fsa/k2
|
9 |
+
icefall: https://github.com/k2-fsa/icefall
|
10 |
+
lhotse: https://github.com/lhotse-speech/lhotse
|
11 |
+
* Install k2 and lhotse, k2 installation guide refers to https://k2.readthedocs.io/en/latest/installation/index.html, lhotse refers to https://lhotse.readthedocs.io/en/latest/getting-started.html#installation. I think the latest version would be ok. And please also install the requirements listed in icefall.
|
12 |
+
* Clone icefall(https://github.com/k2-fsa/icefall) and check to the commit showed above.
|
13 |
+
```
|
14 |
+
git clone https://github.com/k2-fsa/icefall
|
15 |
+
cd icefall
|
16 |
+
```
|
17 |
+
* Preparing data.
|
18 |
+
```
|
19 |
+
cd egs/timit/ASR
|
20 |
+
bash ./prepare.sh
|
21 |
+
```
|
22 |
+
* Training
|
23 |
+
```
|
24 |
+
export CUDA_VISIBLE_DEVICES="0"
|
25 |
+
python tdnn_ligru_ctc/train.py --bucketing-sampler True \
|
26 |
+
--concatenate-cuts False \
|
27 |
+
--max-duration 200 \
|
28 |
+
--world-size 1
|
29 |
+
```
|
30 |
+
## Evaluation results
|
31 |
+
The best decoding results (PER, equals to WER) on TIMIT TEST are listed below, we got this result by averaging models from epoch 9 to 25, the lm_scale is 0.1, the decoding method is `whole-lattice-rescoring`.
|
32 |
+
||TEST|
|
33 |
+
|--|--|
|
34 |
+
|PER|17.66%|
|