This dataset contains Ukrainian expressive single-speaker speech samples prepared for non-commercial research in speech synthesis, speech processing, and related machine learning tasks.

The dataset was prepared as part of research and development work on Ukrainian text-to-speech and digital avatar generation systems. It is intended to support experiments with expressive Ukrainian speech, TTS model training, ASR-based evaluation, and speech processing pipelines.

The dataset focuses on animated-style expressive speech and is provided only for research and educational use.

Dataset overview

Language: Ukrainian
Number of speakers: 1
Speaker label: speaker_01
Speech style: expressive / animated-style speech
Audio format: WAV
Metadata format: LJSpeech-style pipe-separated metadata
Task type: text-to-speech, speech synthesis, speech processing, ASR evaluation
Access type: research-only / non-commercial use

Intended use

This dataset is intended for:

non-commercial research;
educational experiments;
TTS model training and evaluation;
ASR and speech processing experiments;
benchmarking of speech synthesis pipelines;
reproducibility of research results;
experiments with Ukrainian expressive speech synthesis.

Not intended use

This dataset must not be used for:

commercial voice cloning;
paid TTS, dubbing, or voice conversion services;
commercial products, services, or monetized applications;
impersonation of real persons, performers, actors, characters, or public figures;
deceptive, harmful, misleading, defamatory, political, fraudulent, or illegal audio generation;
redistribution, resale, sublicensing, repackaging, or reuploading to other platforms;
removing attribution or changing the stated usage terms.

Dataset structure

dataset/
  wavs/
    s01e01_0000.wav
    s01e01_0001.wav
    s01e01_0002.wav

  metadata.csv
  strict_report.csv
  strict_summary.csv
  README.md
  USAGE_TERMS.md

Metadata format

The main metadata.csv file follows a simple pipe-separated LJSpeech-style format:

id|text|normalized_text
s01e01_0000|Сьогодні важливий день, Геррі.|Сьогодні важливий день, Геррі.

Column description:

id — unique audio segment identifier without the .wav extension;
text — original text transcription;
normalized_text — normalized text used for training and verification.

The corresponding audio file is located in the wavs/ directory and has the same identifier with the .wav extension:

wavs/s01e01_0000.wav

For example, this metadata row:

s01e01_0000|Сьогодні важливий день, Геррі.|Сьогодні важливий день, Геррі.

points to this audio file:

wavs/s01e01_0000.wav

This format is compatible with many LJSpeech-style TTS pipelines where the first column is used as the audio file identifier and the second or third column is used as the target text.

Quality filtering

The dataset was created using a strict automatic filtering pipeline.

Initial candidate segments: 6,252 Accepted segments: 5,463 Rejected segments: 789 Acceptance rate: 87.38%

Filtering criteria:

maximum character error rate: 0.12;
maximum word error rate: 0.30;
minimum target speaker score: 0.50;
minimum speaker margin: 0.04;
head and tail silence validation;
missing audio and metadata validation.

Rejection summary:

rejected by CER threshold: 657;
rejected by WER threshold: 132;
rejected by silence checks: 0;
rejected by duration limits: 0;
rejected by missing files or metadata: 0;
repair failed: 0.

The file strict_report.csv contains per-clip quality information, including duration, CER, WER, ASR hypothesis, target speaker score, speaker margin, source file, segment start time, and segment end time.

The file strict_summary.csv contains the global filtering statistics for the dataset release.

`strict_report.csv`

The strict_report.csv file contains detailed quality-control information for every accepted clip.

Expected columns:

id,text,norm_text,duration_sec,cer,wer,asr_score,asr_hypothesis,target_score,margin,clip_score,speaker,source_json,start,end,bad_head_silence_before_repair,bad_tail_silence_before_repair,wav_path

Column description:

id — unique segment identifier;
text — original text transcription;
norm_text — normalized text used for verification;
duration_sec — audio segment duration in seconds;
cer — character error rate between the target text and ASR hypothesis;
wer — word error rate between the target text and ASR hypothesis;
asr_score — internal ASR confidence or score;
asr_hypothesis — text recognized by the ASR system;
target_score — internal target-speaker score;
margin — speaker confidence margin;
clip_score — optional combined or internal clip score;
speaker — internal speaker label;
source_json — source manifest file;
start — segment start time in the source audio;
end — segment end time in the source audio;
bad_head_silence_before_repair — whether problematic leading silence was detected before repair;
bad_tail_silence_before_repair — whether problematic trailing silence was detected before repair;
wav_path — path to the corresponding WAV file.

For public releases, wav_path should use relative paths such as:

wavs/s01e01_0000.wav

`strict_summary.csv`

The strict_summary.csv file contains global filtering statistics for the dataset release.

Expected columns:

source_total_metadata,kept,min_target_score,min_margin,max_cer,max_wer,bad_head,bad_tail,too_short,too_long,bad_cer,bad_wer,low_target_score,low_margin,missing_manifest,missing_metadata,missing_wav,repair_failed

Column description:

source_total_metadata — total number of candidate segments before strict filtering;
kept — number of accepted segments;
min_target_score — minimum required target-speaker score;
min_margin — minimum required speaker margin;
max_cer — maximum allowed character error rate;
max_wer — maximum allowed word error rate;
bad_head — number of segments rejected due to problematic leading silence;
bad_tail — number of segments rejected due to problematic trailing silence;
too_short — number of segments rejected for being too short;
too_long — number of segments rejected for being too long;
bad_cer — number of segments rejected by CER threshold;
bad_wer — number of segments rejected by WER threshold;
low_target_score — number of segments rejected by target-speaker score;
low_margin — number of segments rejected by speaker margin;
missing_manifest — number of segments rejected due to missing manifest data;
missing_metadata — number of segments rejected due to missing metadata;
missing_wav — number of segments rejected due to missing WAV files;
repair_failed — number of segments rejected because audio repair failed.

Speaker information

The dataset contains one speaker.

For responsible-use and rights-management reasons, the speaker is identified only as:

speaker_01

The dataset must not be used to impersonate any real person, performer, actor, character, or public figure.

License and access

This dataset is distributed for non-commercial research and educational use only.

Access to the dataset may be restricted and provided only to users who agree to the usage terms.

See USAGE_TERMS.md for the full usage policy.

Rights notice

This dataset is made available only for research and educational purposes.

No ownership of the original audiovisual work, character, performance, dubbing, phonogram, or related rights is transferred to dataset users.

Users are responsible for ensuring that their use of the dataset complies with the stated usage terms and applicable law.

Citation

If you use this dataset in academic work, demos, repositories, papers, presentations, or other research materials, please cite it as:

Sukhai Roman. Ukrainian Expressive Single-Speaker TTS Dataset. Version 1.0.

Disclaimer

The dataset is provided “as is”, without warranty of any kind. The dataset maintainer is not responsible for any misuse of the dataset or models trained on it.

Takedown

If you are a rights holder or authorized representative and believe that any material should be removed or restricted, please contact the dataset maintainer.

Downloads last month: 9

Total file size:

703 MB