File size: 2,037 Bytes
9698a33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: corrected_sentence
    dtype: string
  splits:
  - name: train
    num_bytes: 511635794.328
    num_examples: 23113
  - name: validation
    num_bytes: 69413983.807
    num_examples: 2889
  - name: test
    num_bytes: 74404523.28
    num_examples: 2890
  download_size: 673679878
  dataset_size: 655454301.415
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
language:
- fa
pretty_name: Common Voice 17 Persian  ASR-Ready (Cleaned)
task_categories:
  - automatic-speech-recognition
tags:
  - audio
  - speech
  - automatic-speech-recognition
  - asr
  - persian
  - fa
  - farsi
  - iran
  - mozilla
  - common-voice
  - gpt-4o
  - spelling-correction
  - cleaned
  - iranian
size_categories:
  - 10K<n<100K

---
# 🗣️ Common Voice 17 — Persian (Spelling-Corrected Edition)

This is a refined version of the **Persian subset** of Mozilla's Common Voice 17 dataset, specially curated to enhance the performance of ASR (Automatic Speech Recognition) systems in Persian.

## 🛠️ Why this matters
The original dataset contained a significant number of spelling inconsistencies and typographical errors, which negatively impacted transcription accuracy and model alignment.

## ✨ What’s improved
Over **28,000** transcriptions were automatically cleaned using **GPT-4o**, with a focus on preserving the original meaning while correcting orthographic issues.  
The audio files remain unchanged, and all metadata fields are preserved for compatibility.

## 🎯 Use cases
This corrected dataset is ideal for fine-tuning Whisper, Wav2Vec2, or other speech-to-text models in Persian — providing cleaner, more reliable supervision for training and evaluation.

## 🤝 Contribute or collaborate
Feel free to open an issue or submit pull requests if you'd like to contribute further improvements or validations.