patrickvonplaten
commited on
Commit
•
408690c
1
Parent(s):
419a9e3
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Fine-tuning of `wav2vec2-base` on 100h of Librispeech training data. Results on "clean" data are very similar to the ones of the [official model](https://huggingface.co/facebook/wav2vec2-base-100h). However, the result on "other" is significantly worse - the model seems to have overfitting to the "clean" data.
|
2 |
+
|
3 |
+
Model was trained on *librispeech-clean-train.100* with following hyper-parameters:
|
4 |
+
|
5 |
+
- 2 GPUs Titan RTX
|
6 |
+
- Total update steps 13000
|
7 |
+
- Batch size per GPU: 32 corresponding to a *total batch size* of ca. ~1500 seconds
|
8 |
+
- Adam with linear decaying learning rate with 3000 warmup steps
|
9 |
+
- dynamic grouping for batch
|
10 |
+
- fp16
|
11 |
+
- attention_mask was **not** used during training
|
12 |
+
|
13 |
+
Check: https://wandb.ai/patrickvonplaten/huggingface/reports/Project-Dashboard--Vmlldzo1MDI2MTU?accessToken=69z0mrkoxs1msgh71p4nntr9shi6mll8rhtbo6c56yynygw0scp11d8z9o1xd0uk
|
14 |
+
|
15 |
+
*Result (WER)* on Librispeech test:
|
16 |
+
|
17 |
+
| "clean" | "other" |
|
18 |
+
|---|---|
|
19 |
+
| 6.5 | 18.7 |
|