README.md

---
language:
  - ka
tags:
  - asr
  - georgian
  - common-voice
  - speech
  - quality-control
  - dataset-curation
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
pretty_name: "Common Voice 23 — Georgian ASR Quality Control Dataset"
size_categories:
  - 1K<n<10K
citation:
  - "@dataset{giorgadze2025asrge_qc, title={Common Voice 23 — Georgian ASR Quality Control Dataset}, author={Giorgi Giorgadze}, year={2025}, organization={ASR.GE / IRMS}, note={Quality-control annotations generated with custom Georgian ASR models.}}"

---

# 🗣️ Common Voice 23 — Georgian ASR Quality Control Dataset

## Overview  
This dataset was created as part of a **quality control (QC)** process for the **Mozilla Common Voice Corpus 23 (Georgian subset)**.  
The main goal is to identify **corrupted, low-quality, or mislabeled recordings** that might have passed through Common Voice validation but are unsuitable for training or evaluation.

## 🧩 Methodology

- **User Selection**  
  Unique users were selected from the Common Voice 23 Georgian corpus, including only those with **at least 3 valid audio samples**.  

- **Automatic Transcription**  
  Each audio clip was transcribed using **custom Georgian ASR models** developed at [ASR.GE](https://asr.ge).  

- **Error Metrics**  
  The transcriptions were compared to the reference sentences from Common Voice.  
  The following metrics were computed for each sample:
  - **CER (Character Error Rate)**
  - **WER (Word Error Rate)**

  High CER/WER values often correspond to **faulty recordings**, such as noise, silence, or non-speech segments.

## 📊 Dataset Fields

| Column | Description |
|---------|-------------|
| `path` | Relative path to the original Common Voice audio file |
| `reference` | Ground-truth text provided by Common Voice |
| `cer_min` | Minimum Character Error Rate across model predictions |
| `wer_min` | Minimum Word Error Rate across model predictions |

## 🎯 Purpose and Use Cases

This dataset enables:
- **Detection of problematic users or recordings** in the Common Voice Georgian subset.  
- **Improved dataset curation** by filtering out noisy or low-quality samples.  
- **ASR benchmarking** on real-world Georgian data.  

It can be directly linked with **Common Voice 23 (ka)** using the `path` field, which matches the original file structure.

```python
import pandas as pd

qc = pd.read_parquet("preds2_pub.parquet")
print(qc.head())