k2speech/FeruzaSpeech_44100_Hz_tts

NOT AVAILABLE for academic and personal use

FeruzaSpeech_44100_Hz_tts is the same as FeruzaSpeech https://huggingface.co/datasets/k2speech/FeruzaSpeech only the audio is 44100 Hz. It is perfect for Text-To-Speech models

contact: [email protected] or [email protected] to license this dataset This 44100Hz version was never tested in TTS models yet.

---

---

---

---

---

ICNLSPConference: https://www.youtube.com/watch?v=9whj9yzI_s4&ab_channel=ICNLSPConference

Paper: https://arxiv.org/abs/2410.00035

Example test.tsv:

audio	text_latin	text_cyrillic	duration	words_count
test/1076/1076-04.wav	Jahondagi barmoq bilan sanarli mamlakatda koronavirus holati yo‘q. Ulardan biri  Turkmaniston bo‘lsa, keyingisi Shimoliy Koreya.	Жаҳондаги бармоқ билан санарли мамлакатда коронавирус ҳолати йўқ. Улардан бири  Туркманистон бўлса, кейингиси Шимолий Корея.	13.647	15
test/1076/1076-05.wav	Ammo Turkmanistondan olinayotgan xabarlarga ko‘ra, hozir mamlakat Covid-19ning uchinchi, ehtimoliy eng kuchli to‘lqinini boshdan kechirmoqda.	Аммо Туркманистондан олинаётган хабарларга кўра, ҳозир мамлакат Covid-19нинг учинчи, эҳтимолий энг кучли тўлқинини бошдан кечирмоқда.	15.303	15
test/1076/1076-06.wav	O‘zbekistonda mustaqillikka erishilgach, diniy aqidalarni va an’analarni tiklashga ko‘p harakat qilib qelinmoqda.	Ўзбекистонда мустақилликка эришилгач, диний ақидаларни ва анъаналарни тиклашга кўп ҳаракат қилиб қелинмоқда.	12.918	12

Authors

Created by Anna Povey and Katherine Povey.

Dataset Details

• Language(s) (NLP): Uzbek
• License: Other

• Audio files are stored in subfolders (train/, dev/, test/).
• TSV files (train.tsv, dev.tsv, test.tsv) contain metadata with columns:
  • audio: path to the audio file (relative path)
  • text_latin: transcription in Latin script (Uzbek)
  • text_cyrillic: transcription in Cyrillic script (Uzbek)
  • duration: audio length in seconds
  • words_count: number of words in the transcription

FeruzaSpeech_44100_Hz_tts includes "Train", ”Dev” (development) and ”Test” (testing) sets. The corpus contains high-quality, single-channel, 16-bit .wav audio files, available in 44kHz for TTS.

Subset	Duration
Train	52.09h
Dev	2.93h
Test	4.08h

Source Data

Data consists of the Uzbek book, Calikusu, and BBC news articles.

Who are the source language producers?

One female native Uzbek speaker, from Tashkent, Uzbekistan, producing read speech in a perfect recording environment.

Annotation process

Recordings were read from Cyrrilic excerpts of a book and some BBC news articles, which were later converted to Latin using online tools, with some grammatical errors being manually fixed after the use of the conversion calculator. The average recording length was 16 seconds, the minimum length was 4 seconds, and the maximum length is 51 seconds.

Biases

This dataset only contains audio from a single female speaker, so male speakers are not accounted for. The speaker also has a dialect found in Tashkent, Uzbekistan, so other dialects of Uzbek are not considerde in this dataset.

Other Limitations

The data is more formal as it is sourced from a novel and news articles, which doesn't account for casual speech.

Dataset Card Contact

[email protected]

Citation

@article{FeruzaSpeech2024,
  title         = {FeruzaSpeech: A 60 Hour Uzbek Read Speech Corpus with Punctuation, Casing, and Context},
  author        = {Povey, Anna and Povey, Katherine},
  year          = {2024},
  journal       = {arXiv preprint arXiv:2410.00035},
  url           = {https://huggingface.co/datasets/k2speech/FeruzaSpeech}
}