File size: 3,371 Bytes
e3501a9
 
 
 
 
 
 
 
 
 
 
 
bc25934
e3501a9
 
 
 
 
31d14b9
e3501a9
 
 
 
 
 
 
 
 
 
 
 
 
bc25934
e3501a9
9bab4df
 
31d14b9
e3501a9
 
 
 
9bab4df
 
 
 
 
 
e3501a9
 
9bab4df
 
e3501a9
 
 
9bab4df
e3501a9
 
 
9bab4df
 
 
0315b53
e3501a9
 
 
9bab4df
 
e3501a9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9bab4df
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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
---
library_name: transformers
language:
- hu
base_model: openai/whisper-base
tags:
- generated_from_trainer
datasets:
- fleurs
metrics:
- wer
model-index:
- name: Whisper Base Hungarian
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: google/fleurs
      type: fleurs
      config: hu_hu
      split: None
      args: hu_hu
    metrics:
    - name: Wer
      type: wer
      value: 29.48142356294297
---

<!-- 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 Base Hungarian

Ez a modell a finomhangolt változata a [openai/whisper-base](https://huggingface.co./openai/whisper-base) -nek sarpba/big_audio_data_hun_v2 adatkészleten.
Teszteredmények:
On ("google/fleurs", "hu_hu", "test") (képzés közbeni)
- Loss: 0.7999
- Wer Ortho: 33.8788
- Wer: 29.4814

On ("mozilla-foundation/common_voice_17_0", "hu", "test")
- WER: 25.58
- CER: 6.34
- Normalised WER: 21.18
- Normalised CER: 5.31

## Model description

Egyedi adatkészleten magyarra finomhangolt whisper base modell. A jelenleg elérhető legjobb. 
A többi base modelltől nagyságrendellek jobb eredményeket ér el minden adatkészleten!

## Intended uses & limitations

Üzleti cálra a modell a hozzájárulásom nélkül nem használható! Magán célra szabadon felhasználható a whisper esedeti licenszfeltételei szerint!

## Training and evaluation data

A modell hozzávetőleg 1200 óra gondosan válogatott magyar hanganyag alapján készült. A képzés során a tesztek a google/flerus-t használták a fejlődés ellenőrzésére. 
Alatta a mozilla-foundation/common_voice_17_0 eredménye. 

Egyik adatkészlet sem szerepelt a képzési adatok közt, a modell tesztanyaggal nem fertőzött!

## Training procedure

A képzés optimalizációja 3 napig futott a ray[tune] segítségével, a megtalált optimális képzési paraméterekkel a finomhangolás hozzávetőleg 17 órába telt!

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0003
- train_batch_size: 64
- eval_batch_size: 32
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 256
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.05
- training_steps: 8000
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch  | Step | Validation Loss | Wer Ortho | Wer     |
|:-------------:|:------:|:----:|:---------------:|:---------:|:-------:|
| 0.2523        | 0.3770 | 1000 | 0.9703          | 50.8988   | 46.7185 |
| 0.1859        | 0.7539 | 2000 | 0.8605          | 43.4345   | 39.4103 |
| 0.127         | 1.1309 | 3000 | 0.8378          | 40.6107   | 36.0040 |
| 0.1226        | 1.5079 | 4000 | 0.8153          | 38.9189   | 34.1842 |
| 0.1105        | 1.8848 | 5000 | 0.7847          | 36.6018   | 32.1979 |
| 0.0659        | 2.2618 | 6000 | 0.8298          | 35.3752   | 30.6379 |
| 0.0594        | 2.6388 | 7000 | 0.8132          | 34.8255   | 30.2280 |
| 0.0316        | 3.0157 | 8000 | 0.7999          | 33.8788   | 29.4814 |


### Framework versions

- Transformers 4.45.2
- Pytorch 2.3.0+cu121
- Datasets 3.0.1
- Tokenizers 0.20.1