Datasets:
AZ TTS DATA
An Azerbaijani text-to-speech collection with 163,497 audio–text pairs, distributed as sharded Parquet files. It combines portions of three existing datasets with an additional dataset contributed by the maintainer.
Dataset composition
| Portion | Pairs | Source |
|---|---|---|
| Audiobook recordings | 7,513 | Azerbaijani Audiobooks |
| News speech | 29,311 | Azerbaijani News TTS |
| Azerbaijani speech dataset | 10,088 | Azerbaijani TTS Dataset |
| Additional dataset contributed by the maintainer | 116,585 | Maintainer-contributed data |
| Total | 163,497 |
Where applicable, numerals in the transcripts were written out as Azerbaijani words to represent how they are spoken. The Parquet records include embedded audio and transcript text, plus source, speaker, and duration metadata.
Load with 🤗 Datasets
from datasets import load_dataset
dataset = load_dataset("Izahat2/AZ_TTS_DATA", split="train")
example = dataset[0]
print(example["text"])
print(example["speaker_id"], example["duration_s"])
waveform = example["audio"]["array"]
Install the audio-enabled Datasets package with:
pip install "datasets[audio]"
Schema
Each row contains id, audio, text, source, source_id, speaker_id, duration_s, book, chapter, split, orig_audio, speaker_name, speaker_kind, and speaker_id_status.
Speaker IDs are retained when provided by a source. For sources without speaker IDs, unknown_news_tts and unknown_azerbaijani_tts mark missing metadata; they do not mean those recordings belong to one speaker. The audiobook source's original speaker ID (0) is preserved. speaker_id_status describes how the ID was obtained.
Licensing and attribution
This is a mixed-source collection. The linked source datasets have their own terms and attribution requirements. Check those terms before using or redistributing their portions. No single license is asserted here for all files; the maintainer-contributed portion does not replace the source datasets' terms.
- Downloads last month
- 53