README.md

metadata

language:
  - lg
  - sw
license: cc0-1.0
multilinguality:
  - multilingual
task_categories:
  - text-to-speech
  - automatic-speech-recognition
pretty_name: Luganda-Swahili Speaker-Clustered TTS Dataset
size_categories:
  - 10K<n<100K
tags:
  - tts
  - speech
  - luganda
  - swahili
  - kiswahili
  - low-resource
  - african-languages

Luganda-Swahili Speaker-Clustered TTS Dataset

Dataset Summary

A cleaned, speaker-labeled text-to-speech (TTS) dataset covering Luganda and Kiswahili, derived from the Luganda-Swahili Speech for Text-to-Speech Synthesis Kaggle dataset. The source data contains recordings from 6 speakers per language, but speaker identity was not labeled in the original release. Speaker labels in this version were recovered via unsupervised audio clustering (6 clusters per language), and the audio was filtered for corrupt files before merging into a single train/validation split.

• Languages: Luganda (lg), Kiswahili / Swahili (sw)
• Total utterances: 28,349 (after cleaning)
• Total duration: ~32.16 hours (13.10h Kiswahili + 19.06h Luganda)
• Speakers: 6 per language (12 total), speaker identity inferred via clustering — not ground-truth speaker metadata
• License: CC0 1.0 Public Domain

Supported Tasks

• Text-to-Speech (TTS): text / english_transcript → audio
• Automatic Speech Recognition (ASR): audio → text
• Speaker-conditioned TTS / voice cloning: using the clustered speaker_id field

Dataset Structure

Data Instances

Each example contains a single audio clip paired with its transcript in the source language, an English translation, a language tag, and a clustered speaker ID.

{
  "audio": {"array": [...], "sampling_rate": 16000, "path": "..."},
  "text": "...",
  "english_transcript": "...",
  "language": "lug",  # or "swa"
  "speaker_id": "lug_spk5"
}

Data Fields

Field	Type	Description
audio	Audio	The speech recording
text	string	Transcript in the source language (Luganda or Kiswahili)
english_transcript	string	English translation of the transcript
language	string	Source language code (lug or swa)
speaker_id	string	Clustered speaker label (e.g. lug_spk1–lug_spk6, swa_spk1–swa_spk6); assigned via unsupervised clustering, not original metadata

Data Splits

The dataset was split 90/10 into train/validation, stratified per speaker so every speaker is proportionally represented in both splits (rather than a single global random split, which risked dropping low-count speakers entirely from one side).

Split	Utterances	%
train	25,514	90.0%
validation	2,835	10.0%
Total	28,349	100%

Per-speaker distribution:

Speaker	Train	Validation
lug_spk1	2,172	241
lug_spk2	1,910	212
lug_spk3	3,078	342
lug_spk4	1,867	208
lug_spk5	3,592	399
lug_spk6	2,862	318
swa_spk1	969	108
swa_spk2	1,412	157
swa_spk3	1,785	198
swa_spk4	1,525	169
swa_spk5	2,709	301
swa_spk6	1,633	182

Dataset Creation

Source Data

Raw audio and transcripts were sourced from the Luganda-Swahili Speech for Text-to-Speech Synthesis dataset on Kaggle, released under CC0 1.0 Public Domain.

Speaker Clustering

The original dataset does not label which of the 6 speakers per language produced each clip. Speaker identity was recovered by:

1. Extracting audio features from each clip.
2. Clustering clips per language into 6 groups (k=6) to approximate the 6 known speakers.

Resulting cluster sizes:

Luganda (17,201 clips clustered into 6 speakers):

Speaker	Clips
lug_spk5	3,991
lug_spk3	3,420
lug_spk6	3,180
lug_spk1	2,413
lug_spk2	2,122
lug_spk4	2,075

Kiswahili (11,149 clips clustered into 6 speakers):

Speaker	Clips
swa_spk5	3,010
swa_spk3	1,983
swa_spk6	1,816
swa_spk4	1,694
swa_spk2	1,569
swa_spk1	1,077

Note: these speaker labels are cluster assignments, not verified ground-truth speaker identities. They should be treated as a best-effort approximation useful for speaker-conditioned modeling, not as authoritative metadata.

Data Cleaning

Corrupt/empty audio files were filtered out prior to clustering and merging:

Language	Before	After	Removed
Kiswahili	11,149	11,148	1
Luganda	17,201	17,201	0

Audio Statistics

	Kiswahili	Luganda
Average duration	4.23s	3.99s
Median duration	4.14s	3.78s
Min duration	1.05s	1.17s
Max duration	8.73s	9.72s
Total duration	13.10h	19.06h

Licensing Information

This dataset is released under CC0 1.0 Universal (Public Domain Dedication), matching the license of the source Kaggle dataset.

Citation

If you use this dataset, please cite the original source:

@misc{lugswa_tts_kaggle,
  title  = {Luganda-Swahili Speech for Text-to-Speech Synthesis},
  author = {Dumlao, Jocelyn},
  year   = {2024},
  url    = {https://www.kaggle.com/datasets/jocelyndumlao/luganda-swahili-speech-for-text-to-speechsynthesis/data}
}

Acknowledgements

Speaker clustering, corrupt-file filtering, and the stratified train/validation split were performed as part of dataset preparation work under r-labs.