Update README.md
Browse files
README.md
CHANGED
|
@@ -1,5 +1,42 @@
|
|
| 1 |
-
|
|
|
|
|
|
|
| 2 |
|
| 3 |
-
|
|
|
|
| 4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Conformer-CTC model for speech recognition
|
| 2 |
+
The model was trained on full [LibriSpeech](http://openslr.org/12/) with the scripts in [icefall](https://github.com/k2-fsa/icefall).
|
| 3 |
+
See (https://github.com/k2-fsa/icefall/pull/13) for more details of this model.
|
| 4 |
|
| 5 |
+
## How to use
|
| 6 |
+
See (https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/conformer_ctc/README.md)
|
| 7 |
|
| 8 |
+
## Training procedure
|
| 9 |
+
The version of the mainly repositories are list below.
|
| 10 |
+
k2: https://github.com/k2-fsa/k2/commit/81cec9ec736d2c603ad75d933bb3e3a3706fb0dd
|
| 11 |
+
icefall: https://github.com/k2-fsa/icefall/commit/ef233486ae6d21bacb940de45efb35d0c334605c
|
| 12 |
+
lhotse: https://github.com/lhotse-speech/lhotse/commit/5dfe0f4c02b1334ebb7db6d67e1141fe406ca76b
|
| 13 |
|
| 14 |
+
* 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. It is better to use the given version above, but I think the latest version would be ok. And also install the requirements listed in icefall.
|
| 15 |
+
|
| 16 |
+
* Clone icefall(https://github.com/k2-fsa/icefall) and check to the commit showed above.
|
| 17 |
+
```
|
| 18 |
+
git clone https://github.com/k2-fsa/icefall
|
| 19 |
+
cd icefall
|
| 20 |
+
git checkout ef233486
|
| 21 |
+
```
|
| 22 |
+
* Preparing data.
|
| 23 |
+
```
|
| 24 |
+
cd egs/librispeech/ASR
|
| 25 |
+
bash ./prepare.sh
|
| 26 |
+
```
|
| 27 |
+
* Training
|
| 28 |
+
```
|
| 29 |
+
export CUDA_VISIBLE_DEVICES="0,1,2,3"
|
| 30 |
+
python conformer_ctc/train.py --bucketing-sampler True \
|
| 31 |
+
--concatenate-cuts False \
|
| 32 |
+
--max-duration 200 \
|
| 33 |
+
--full-libri True \
|
| 34 |
+
--world-size 4
|
| 35 |
+
```
|
| 36 |
+
|
| 37 |
+
## Evaluation results
|
| 38 |
+
The best decoding results (WERs) on LibriSpeech test-clean and test-other are listed below, we got this results by averaging models from epoch 15 to 34.
|
| 39 |
+
|
| 40 |
+
||test-clean|test-other|
|
| 41 |
+
|--|--|--|
|
| 42 |
+
|WER|2.57%|5.94%|
|