File size: 3,092 Bytes
060ea69
675d607
060ea69
 
 
675d607
060ea69
a6047ae
060ea69
 
 
675d607
060ea69
 
 
 
 
675d607
a6047ae
060ea69
 
 
 
 
 
a6047ae
060ea69
 
 
 
 
675d607
060ea69
675d607
060ea69
a6047ae
 
060ea69
675d607
060ea69
675d607
060ea69
675d607
060ea69
675d607
060ea69
675d607
060ea69
675d607
060ea69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
675d607
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
---
language: pt
license: apache-2.0
tags:
- generated_from_trainer
- whisper-event
datasets:
- mozilla-foundation/common_voice_11_0
metrics:
- wer
model-index:
- name: openai/whisper-medium
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: mozilla-foundation/common_voice_11_0
      type: mozilla-foundation/common_voice_11_0
      config: pt
      split: test
      args: pt
    metrics:
    - name: Wer
      type: wer
      value: 6.598745817992301
---

<!-- 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. -->

# Portuguese Medium Whisper

This model is a fine-tuned version of [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) on the common_voice_11_0 dataset.
It achieves the following results on the evaluation set:
- Loss: 0.2628
- Wer: 6.5987

## Blog post

All information about this model in this blog post: [Speech-to-Text & IA | Transcreva qualquer áudio para o português com o Whisper (OpenAI)... sem nenhum custo!](https://medium.com/@pierre_guillou/speech-to-text-ia-transcreva-qualquer-%C3%A1udio-para-o-portugu%C3%AAs-com-o-whisper-openai-sem-ad0c17384681).

## New SOTA

The Normalized WER in the [OpenAI Whisper article](https://cdn.openai.com/papers/whisper.pdf) with the [Common Voice 9.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_9_0) test dataset is 8.1. 

As this test dataset is similar to the [Common Voice 11.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0) test dataset used to evaluate our model (WER and WER Norm), it means that **our Portuguese Medium Whisper is better than the [Medium Whisper](https://huggingface.co/openai/whisper-medium) model at transcribing audios Portuguese in text** (and even better than the [Whisper Large](https://huggingface.co/openai/whisper-large) that has a WER Norm of 7.1!).

![OpenAI results with Whisper Medium and Test dataset of Commons Voice 9.0](https://huggingface.co/pierreguillou/whisper-medium-portuguese/resolve/main/whisper_medium_portuguese_wer_commonvoice9.png)

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 9e-06
- train_batch_size: 32
- 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: 6000
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step | Validation Loss | Wer    |
|:-------------:|:-----:|:----:|:---------------:|:------:|
| 0.0333        | 2.07  | 1500 | 0.2073          | 6.9770 |
| 0.0061        | 5.05  | 3000 | 0.2628          | 6.5987 |
| 0.0007        | 8.03  | 4500 | 0.2960          | 6.6979 |
| 0.0004        | 11.0  | 6000 | 0.3212          | 6.6794 |


### Framework versions

- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2