NMikka/Common-Voice-Geo-Cleaned

Common Voice Georgian — Cleaned for TTS/STT

A high-quality subset of Mozilla Common Voice Georgian cleaned and filtered specifically for text-to-speech fine-tuning.

Dataset Summary

Splits

Quality Pipeline

The dataset was cleaned from ~71K raw Common Voice recordings through a 6-stage pipeline:

1. 1.Standardize — Resample to 24 kHz mono, normalize loudness to −23 LUFS, filter duration to [0.5s, 30s]
2. 2.Enhance — VoiceFixer audio restoration + Sox spectral noise subtraction
3. 3.NISQA Filter — NISQA MOS ≥ 3.0 (neural speech quality assessment)
4. 4.Duration Outlier — IQR-based character duration filter (removes misaligned/rushed/slow speech)
5. 5.Transcript Verify — Round-trip ASR (Meta Omnilingual 7B, 1.9% CER on Georgian) with CER ≤ 0.20 threshold
6. 6.Speaker Select — Keep speakers with ≥ 1800 seconds total audio

Fields

Speaker Distribution

Statistics

• —Duration: min 2.4s, mean 5.9s, max 10.6s

Usage

from datasets import load_dataset

ds = load_dataset("NMikka/Common-Voice-Geo-Cleaned")

# Training
for sample in ds["train"]:
    print(sample["text"], sample["duration"])

# Validation
for sample in ds["eval"]:
    print(sample["text"])

# Best speaker references (for TTS inference/voice cloning)
for sample in ds["test"]:
    print(sample["text"], sample["speaker_id"])

Citation

If you use this dataset, please cite Mozilla Common Voice:

@inproceedings{ardila2020common,
  title={Common Voice: A Massively-Multilingual Speech Corpus},
  author={Ardila, Rosana and others},
  booktitle={LREC},
  year={2020}
}