File size: 2,000 Bytes
291bc98
254de90
 
291bc98
254de90
 
 
 
 
 
 
 
 
291bc98
254de90
 
 
 
 
 
 
 
 
 
 
 
 
de22c9c
254de90
 
 
ca6ac25
254de90
 
 
de22c9c
254de90
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
---
language:
- tr
license: apache-2.0
base_model: openai/whisper-medium
tags:
- hf-asr-leaderboard
- generated_from_trainer
metrics:
- wer
model-index:
- name: Whisper Medium Tr - denysdios
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# Whisper Medium Tr - denysdios

This model is a fine-tuned version of [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) on the Common Voice 13.0 & Fleurs dataset.
It achieves the following results on the evaluation set:
- Loss: 0.1618
- Wer: 14.3825

## Model description

The model took about nine hours to train on a single A100 GPU.

## Intended uses & limitations

Absolutely no restrictions additional to whisper models. Increasing the Turkish labeled data in whisper, which was 4333/690k (0.0063), was the primary objective. There are just 49.945 hours of data in the fine-tuning dataset, or about 1.1% of the Turkish dataset that has already been trained. 

## Training and evaluation data

Processing...

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 4000
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step | Validation Loss | Wer     |
|:-------------:|:-----:|:----:|:---------------:|:-------:|
| 0.1803        | 0.36  | 1000 | 0.2089          | 18.6326 |
| 0.1428        | 0.71  | 2000 | 0.1821          | 16.3912 |
| 0.0535        | 1.07  | 3000 | 0.1693          | 14.9132 |
| 0.0491        | 1.43  | 4000 | 0.1618          | 14.3825 |


### Framework versions

- Transformers 4.38.1
- Pytorch 2.1.0+cu121
- Datasets 2.17.1
- Tokenizers 0.15.2