Hugging Face's logo

Datasets:
hostbot77
/
Speech_to_fine_tuning

Tasks:
Automatic Speech Recognition
Text-to-Audio
Modalities:
Audio
Text
Formats:
parquet
Languages:
Uzbek
Size:
10K - 100K
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
Community
1
Speech_to_fine_tuning / README.md
hostbot77's picture
hostbot77
Duplicate from nickoo004/FeruzaSpeech_to_fine_tuning
f13dc12
preview code
|
raw
history blame contribute delete
3.78 kB
metadata 
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: text
      dtype: string
    - name: duration
      dtype: float64
    - name: token_count
      dtype: int64
  splits:
    - name: train
      num_bytes: 5374523458.74
      num_examples: 11444
    - name: dev
      num_bytes: 338070338
      num_examples: 648
    - name: test
      num_bytes: 470470140
      num_examples: 899
  download_size: 13334454656
  dataset_size: 6183063936.74
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: dev
        path: data/dev-*
      - split: test
        path: data/test-*
license: apache-2.0
task_categories:
  - automatic-speech-recognition
  - text-to-audio
language:
  - uz
pretty_name: FeruzaSpeech
size_categories:
  - 10K<n<100K

FeruzaSpeech_to_fine_tuning

A speech corpus of ⏱️ ~59.1 total hours of Uzbek audio paired with Latin‑script transcripts, intended for fine‑tuning ASR / speech‑to‑text models.

Dataset Details

Dataset Description

This dataset contains recordings of native Uzbek speakers reading a mix of classical literature excerpts and school‑level writing prompts:

  • 001: Choliqushi (a novel by Rashod Nuri Guntekin, trans. by Mirzakalon Ismoiliy; first pub. Sept 1900).
  • 002: Excerpts from Uzbek secondary‑school essays (“To‘rtinchi sinfda edim…”, “Hayotdagi ilk xotiralaringizni yozing…”).
  • .....

Each line in text_latin.txt is of the form:

//

We strip the filename prefix in preprocessing so that the text field contains only the spoken words.

Dataset Statistics

Split # Examples Total Size Approx. Duration
train 11 444 5.37 GB 52.09 hours
dev 648 0.34 GB 2.93 hours
test 899 0.47 GB 4.08 hours

Dataset Creation

Curation Rationale

We aim to provide a high‑quality, publicly available Uzbek ASR dataset combining both literary and educational domains to improve model robustness.

Source Data

  • Audio recorded in a quiet home‑studio environment, 16 kHz mono WAV, 16‑bit PCM.
  • Transcripts created from existing texts (classical novels, school writing prompts).

Who Are the Source Data Producers?

  • Recordings & Transcriptions by: k2speech/FeruzaSpeech
  • Translators / Editors: Nickoo 004

Uses

Direct Use

Fine‑tuning or evaluating speech‑to‑text/ASR models for Uzbek. It’s also suitable for speech processing research (voice activity detection, speaker diarization, etc.).

Out‑of‑Scope Use

  • Speaker identification / sensitive demographic inference.
  • Real‑time speech generation.

Supported Tasks and Leaderboards

  • Task: Automatic Speech Recognition

Dataset Structure

Each example has the following fields:

  • audio: an Audio object (array + sampling_rate)
  • text: Latin‑script transcript, cleaned of filename tokens
  • duration: audio length in seconds
  • token_count: length of the transcript in raw word‑piece tokens

Distribution

Who Maintains This Dataset

Citation

If you use this dataset, please cite:

@misc{feruzaspeech2025,
  title        = {FeruzaSpeech\_to\_fine\_tuning: An Uzbek ASR Dataset},
  author       = {Nickoo\, 004},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/nickoo004/FeruzaSpeech_to_fine_tuning}},
  license      = {Apache 2.0}
}