bambara-synthetic-audio

102,310 utterances of synthetic Bambara speech, generated by a text-to-speech model over Bambara text. Every clip is machine-generated; no human voice is recorded here.

Load

from datasets import load_dataset

# Enhanced set, with per-clip quality scores
semi = load_dataset("djelia/bambara-synthetic-audio", "semi-clean", split="train")

# Larger generation set, no quality scores
v2 = load_dataset("djelia/bambara-synthetic-audio", "tts_v2", split="train")

Note the config names: semi-clean uses a hyphen, tts_v2 an underscore.

Configs

Config	Split	Rows	Audio	Quality scores	Size
semi-clean	train	36,685	69.18 h	yes	3.03 GB
tts_v2	train	65,625	≈115 h over 59,660 rows measured	no	5.51 GB

Fields

Field	Type	Notes
text	string	Bambara text given to the TTS model
speaker_description	string	Conditioning prompt; 4 distinct voices (Sekou, Seydou, Mariam, Moussa)
audio	Audio	The generated waveform
duration	float64	Seconds; mean 6.8 s, max 45.9 s
created_at	timestamp	When the clip was generated
stoi, pesq, si_sdr	float64	Objective speech-quality scores, semi-clean only

tts_v2 has the same fields minus the three quality columns.

Notes

The audio feature declares no sampling rate; read row["audio"]["sampling_rate"] rather than assuming one.

Quality is long-tailed in semi-clean — si_sdr runs down to −11.4 dB and stoi to 0.41 — so filter if you need clean clips:

good = semi.filter(lambda row: row["pesq"] >= 3.0 and row["si_sdr"] >= 15.0)

The two configs are not disjoint; they come from one generation campaign, so measure the overlap on (text, created_at) before concatenating them.

Stream with streaming=True to avoid pulling 8.54 GB.