BlueRaccoon
commited on
Commit
•
a4182c5
1
Parent(s):
34efa15
add some more info to the model card
Browse files
README.md
CHANGED
@@ -40,19 +40,24 @@ model-index:
|
|
40 |
name: WER
|
41 |
---
|
42 |
|
43 |
-
<!--
|
44 |
-
|
45 |
|
46 |
# Whisper Small Uzbek
|
47 |
|
48 |
This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) trained on the mozilla-foundation/common_voice_11_0 uz and google/fleurs uz_uz datasets, and evaluated on the mozilla-foundation/common_voice_11_0 uz dataset.
|
49 |
-
It achieves the following results on the evaluation set:
|
50 |
- Loss: 0.3872
|
51 |
- Wer: 23.6509
|
|
|
|
|
52 |
|
53 |
## Model description
|
54 |
|
55 |
-
|
|
|
|
|
|
|
56 |
|
57 |
## Intended uses & limitations
|
58 |
|
@@ -60,10 +65,14 @@ More information needed
|
|
60 |
|
61 |
## Training and evaluation data
|
62 |
|
63 |
-
|
|
|
64 |
|
65 |
## Training procedure
|
66 |
|
|
|
|
|
|
|
67 |
### Training hyperparameters
|
68 |
|
69 |
The following hyperparameters were used during training:
|
|
|
40 |
name: WER
|
41 |
---
|
42 |
|
43 |
+
<!-- Disclaimer: I've never written a model card before. I'm probably not correctly following standard practices on how they should be written.
|
44 |
+
I'm new to this. I'm sorry -->
|
45 |
|
46 |
# Whisper Small Uzbek
|
47 |
|
48 |
This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) trained on the mozilla-foundation/common_voice_11_0 uz and google/fleurs uz_uz datasets, and evaluated on the mozilla-foundation/common_voice_11_0 uz dataset.
|
49 |
+
It achieves the following results on the common_voice_11_0 evaluation set:
|
50 |
- Loss: 0.3872
|
51 |
- Wer: 23.6509
|
52 |
+
It achieves the following results on the FLEURS evaluation set:
|
53 |
+
- Wer: 47.15
|
54 |
|
55 |
## Model description
|
56 |
|
57 |
+
This model was created as part of the Whisper fine-tune sprint event.
|
58 |
+
Based on eval, this model achieves a WER of 23.6509 against the Common Voice 11 dataset and 47.15 against the FLEURS dataset.
|
59 |
+
This is a significant improvement over the reported WER of 90.2 recorded on the [Whisper article](https://cdn.openai.com/papers/whisper.pdf):
|
60 |
+
![A part of Table 13 from the paper "Robust Speech Recognition via Large-Scale Weak Supervision", which shows the WER achieved by the Whisper model under the FLEURS dataset. Highlighted is the best score it achieved under for the Uzbek language, which was 90.2.](https://huggingface.co/BlueRaccoon/whisper-small-uz/resolve/main/uzbektable13.png)
|
61 |
|
62 |
## Intended uses & limitations
|
63 |
|
|
|
65 |
|
66 |
## Training and evaluation data
|
67 |
|
68 |
+
Training was performed using the train and evaluation splits from [Mozilla's Common Voice 11](https://huggingface.co/mozilla-foundation/common_voice_11_0) and [Google's FLEURS](https://huggingface.co/google/fleurs) datasets.
|
69 |
+
Testing was performed using the test splits from the same datasets.
|
70 |
|
71 |
## Training procedure
|
72 |
|
73 |
+
Training and CV11 testing was performed using a modified version of the [run_speech_recognition_seq2seq_streaming.py](https://github.com/kamfonas/whisper-fine-tuning-event/blob/e0377f55004667f18b37215d11bf0e54f5bda463/run_speech_recognition_seq2seq_streaming.py) script by farsipal, which enabled training on multiple datasets in a convenient way.
|
74 |
+
FLEURS testing was performed using the standard [run_eval_whisper_streaming.py](https://github.com/huggingface/community-events/blob/main/whisper-fine-tuning-event/run_eval_whisper_streaming.py) script.
|
75 |
+
|
76 |
### Training hyperparameters
|
77 |
|
78 |
The following hyperparameters were used during training:
|