Dhivehi Audio Dataset 1
A quality-filtered Dhivehi speech dataset with three subsets (bronze, silver, gold), each representing a progressively stricter quality threshold.
Subsets
| Subset | MOS | CTC gc | WER | Train | Test |
|---|---|---|---|---|---|
| bronze | ≥3.0 | ≥0.70 | — | 224,321 | 24,925 |
| silver | ≥3.0 | ≥0.70 | ≤0.30 | 131,389 | 14,599 |
| gold | ≥3.5 | ≥0.75 | ≤0.20 | 15,535 | 1,727 |
- MOS — Mean Opinion Score (1–4), subjective listening quality rating
- CTC gc — CTC forced-alignment geo-confidence (0–1), measures pronunciation accuracy
- WER — Word Error Rate from ASR transcription vs reference text (lower is better)
Each subset is a strict superset of the next: gold ⊂ silver ⊂ bronze.
Fields
| Field | Type | Description |
|---|---|---|
| audio | Audio (16 kHz, mono) | Speech audio |
| sentence | string | Dhivehi text (numbers spelled out in Thaana) |
| gender | string | Speaker gender (male / female) |
Usage
from datasets import load_dataset
ds = load_dataset("alakxender/dhivehi-audios-ds1", "silver")
print(ds["train"][0])
- Downloads last month
- 97