Denube-final

Precomputed SpeechT5 TTS training tensors for Bambara: 30,765 examples of (character ids → 80-bin log-mel spectrogram) conditioned on a 512-dimensional x-vector speaker embedding. There is no audio column — the rows are features, not waveforms.

Load

from datasets import load_dataset

train = load_dataset("djelia/Denube-final", split="train")
val = load_dataset("djelia/Denube-final", split="eval")     # "eval", not "validation"
test = load_dataset("djelia/Denube-final", split="test")

row = train[0]
print(len(row["speaker_embeddings"]))              # 512
print(len(row["labels"]), len(row["labels"][0]))   # n_frames, 80

Config	Split	Rows	Mel frames (mean)
default	train	28,641	200.1
default	eval	1,508	202.6
default	test	616	209.4

Fields

Field	Type	Notes
input_ids	list	Character-level token ids, len(text) + 2 per row
labels	list<list>	[n_frames, 80] log-mel spectrogram, 27–3,411 frames
speaker_embeddings	list	512-dim x-vector, present on every row
speaker_id	int32	69 distinct values in [-1, 67]
text	string	Source Bambara text

Notes

input_ids and labels are already a processor's output, so do not run a SpeechT5Processor over them again. Lengths are ragged — use a collator that pads and masks input_ids and labels to the batch maximum.

Text is ASCII-folded: ɲ is written gn, ɛ/ɔ as e/o, and no row contains ɛ ɔ ɲ ŋ. A model trained here expects folded input at inference time. For the same corpus with audio and standard orthography, see djelia/danube.