Update README.md
Browse files
README.md
CHANGED
@@ -7,4 +7,27 @@ language:
|
|
7 |
---
|
8 |
# Whisper Tiny RixVox Swedish
|
9 |
|
10 |
-
This is a Whisper-tiny finetuned for Swedish using
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
---
|
8 |
# Whisper Tiny RixVox Swedish
|
9 |
|
10 |
+
This is a [Whisper tiny](https://huggingface.co/openai/whisper-tiny) finetuned for Swedish using
|
11 |
+
the [RixVox](https://huggingface.co/datasets/KBLab/rixvox) dataset.
|
12 |
+
|
13 |
+
## Evaluation
|
14 |
+
|
15 |
+
### [Common Voice 11](#):
|
16 |
+
* WER: XYZ
|
17 |
+
* WER (normalized): XYZ
|
18 |
+
|
19 |
+
## Training
|
20 |
+
|
21 |
+
Training was done using Huggingface and Deepspeed with ZeRO stage 2.
|
22 |
+
|
23 |
+
* learning rate: 1e-5
|
24 |
+
* optimizer: CPUAdamW (Deepspeed)
|
25 |
+
* lr scheduler: linear
|
26 |
+
* warmup steps: 500
|
27 |
+
* per device batch size: 32
|
28 |
+
* GPUs: 8 x NVIDIA A100 40GB
|
29 |
+
* total batch size: 160
|
30 |
+
* steps: 10000
|
31 |
+
* lowercase: no
|
32 |
+
* fp16
|
33 |
+
* entire decoder was frozen
|