license: cc-by-3.0
language:
- uz
task_categories:
- automatic-speech-recognition
- text-to-speech
pretty_name: YDS
size_categories:
- 100K<n<1M
dataset_info:
features:
- name: sample_id
dtype: string
- name: audio
dtype: audio
- name: text
dtype: string
- name: language
dtype: string
- name: duration_sec
dtype: float32
- name: sample_rate
dtype: int32
- name: word_timestamps_json
dtype: string
- name: ctc_mean_score
dtype: float32
- name: speaker_overlap_sec
dtype: float32
- name: recording_id
dtype: string
- name: source_start_ms
dtype: int64
- name: source_end_ms
dtype: int64
configs:
- config_name: uz
data_files:
- split: train
path: data/uz/*.parquet
YDS
YDS is a filtered Uzbek speech and transcript dataset. This release contains 196,412 clips (240.62 hours). It is public and requires manual approval for access.
Data
Each row has embedded mono FLAC audio, an Uzbek transcript, word-level CTC timings in JSON, duration, sample rate, alignment score, speaker overlap, and source recording/time identifiers. There is one uz configuration and one train split. Speaker labels across clips are not established identities.
Load with the Hugging Face datasets library using load_dataset("instinct1912/yds", "uz", split="train").
Processing and quality filters
Source segments were restored on a GPU pod, transcribed in Uzbek with Gemini 3.8 Flash, and accepted only when intelligible. One accepted transcription per clip was retained. Accepted text was CTC-aligned to audio; clips with failed alignment or audio/text checks were excluded. Pyannote diarization then retained clips assessed as one speaker with no more than 0.10 seconds of overlapping speech. Passing clips were trimmed around the first and last aligned words with up to 120 ms before and 180 ms after, then checked again for duration, silence, clipping, peak, and non-finite samples. This dataset's final check permits edge silence because source segments were VAD processed.
Of 198,255 speaker-passed clips, 196,412 passed final trim and audio QA; 1,843 were excluded. Automated checks reduce errors but do not guarantee every transcription or speaker assignment is correct. Recordings may contain background noise and varied recording conditions. No evaluation split has been curated.
Source and license
This is a processed subset of ESPnet YODAS v3. The source dataset is distributed under CC BY 3.0; attribution to ESPnet and the original dataset is required. This release does not contain the source dataset's raw archives or unselected recordings.