Commit
•
9fe5087
1
Parent(s):
bb7e82a
Create k2_zipformer2_english_v1_small.md
Browse files
k2_zipformer2_english_v1_small.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# k2_zipformer2_english_v1
|
| 2 |
+
|
| 3 |
+
- Zipformer2 recipe derived from : https://github.com/k2-fsa/icefall/tree/master/egs/librispeech/ASR/zipformer
|
| 4 |
+
|
| 5 |
+
- Training data : CommonVoice, VoxPopuli (speed-perturb 3x):
|
| 6 |
+
- cca 1500 + 3x500 hours of training data
|
| 7 |
+
|
| 8 |
+
- Output text/symbols include:
|
| 9 |
+
- TrueCase capitalization
|
| 10 |
+
- punctuation `[,.?!]` as standalone tokens
|
| 11 |
+
|
| 12 |
+
## Config:
|
| 13 |
+
```
|
| 14 |
+
--num-epochs 20 \
|
| 15 |
+
--base-lr 0.04 \
|
| 16 |
+
\
|
| 17 |
+
--causal 1 \
|
| 18 |
+
--use-transducer 1 \
|
| 19 |
+
--use-ctc 0 \
|
| 20 |
+
\
|
| 21 |
+
--num-encoder-layers 2,2,2,2,2,2 \
|
| 22 |
+
--feedforward-dim 512,768,768,768,768,768 \
|
| 23 |
+
--encoder-dim 192,256,256,256,256,256 \
|
| 24 |
+
--encoder-unmasked-dim 192,192,192,192,192,192 \
|
| 25 |
+
```
|
| 26 |
+
|
| 27 |
+
## Results
|
| 28 |
+
|
| 29 |
+
| ID | System | cv-dev | cv-test | vp-dev | vp-test | Comment |
|
| 30 |
+
|---|-------------------------|--------|---------|--------|---------|-------------------------------|
|
| 31 |
+
| A | small (24M) | 18.57 | 21.98 | 13.66 | 13.26 | ep20,avg4 |
|
| 32 |
+
|
| 33 |
+
- non-streaming results from [decode.py](https://github.com/BUTSpeechFIT/k2_streaming_training/blob/main/training/zipformer/decode.py)
|
| 34 |
+
- cv = CommonVoice, vp = VoxPopuli
|
| 35 |
+
- exported the model averaging : ep20,avg4
|
| 36 |
+
|
| 37 |
+
## Note
|
| 38 |
+
- Not the best possible results, this model is for integration tests
|