Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is reviewed manually. Please tell us who you are and how you plan to use the data.
Log in or Sign Up to review the conditions and access this dataset content.
Mooré Speech Corpus
Mooré (mos) speech with transcripts, curated by Burkimbia. It combines two datasets:
subset |
Source | Rows | Hours |
|---|---|---|---|
cleaned |
burkimbia/speech-cleaned: clips that passed the curation filters |
89,055 | 126.2 |
rejected_aligned |
burkimbia/speech-rejected-aligned: long rejected clips, forced-aligned and cut into segments |
1,129 | 5.0 |
| Total | 90,184 | 131.2 |
By source: BIBLE.com 62.5 h (56,992 rows), JW 65.4 h (29,070), contes 1.9 h (2,579), proverbes 1.0 h (965), devinettes 0.4 h (578).
Fields
| Field | Description |
|---|---|
id |
Unique row id |
recording_id |
Id of the recording the clip was cut from; clips of one recording share it, across both subsets |
audio |
16 kHz audio |
text |
Mooré transcript |
duration |
Length in seconds |
source |
Where the recording comes from: BIBLE.com, JW, contes, proverbes, devinettes |
subset |
cleaned or rejected_aligned |
speakers, gender |
Speaker label and gender |
duplicate, clipped, silence_ratio, cps, word_count, detected_lang |
Cleaning metrics, for all rows |
speakers_all, start_time, end_time, align_score |
rejected_aligned only (null for cleaned): all speaker labels of the segment's copies, position in the source clip, alignment confidence |
The rejected_aligned rows went through the same cleaning filter as cleaned. 14 of them fail its repetition check, but they are Bible genealogies and name lists, read correctly, so they are kept.
Notes
- Use
recording_idto keep clips of one recording in the same split. duplicate=Truemarks 13,032 rows incleanedthat repeat another recording. Filter them out if you need unique audio.- 98% of the audio is male speech (129.2 h male, 2.0 h female), and 97% is read from Bible texts (JW and BIBLE.com).
Usage
from datasets import load_dataset
ds = load_dataset("burkimbia/moore-speech-corpus", split="train")
ds = ds.filter(lambda d: d is not True, input_columns="duplicate")
License
CC BY-NC 4.0: non-commercial use, with attribution.
- Downloads last month
- 375