Turkish Speech Dataset With 41,427 Audio Segments

41,427 audio segments from 88 source datasets, merged into a single collection for Turkish speech. The dataset includes 222 speakers and provides transcriptions alongside emotion labels such as neutral, angry, sad, and happy.

Use Cases

• Train a text-to-speech model using the audio and text fields for Turkish language synthesis.
• Build an emotion recognition classifier for speech by training on the audio and emotion labels.
• Perform speaker verification or identification tasks using the audio and speaker_id fields across 222 speakers.
• Fine-tune an automatic speech recognition model on Turkish language using the paired audio and text data.

Strengths

• Contains 41,427 audio segments, providing a substantial sample size for model training.
• Aggregates data from 88 original datasets, increasing diversity of sources and recording conditions.
• Includes four distinct emotion labels (neutral, angry, sad, happy) for each audio segment.
• Covers 222 unique speakers, enabling analysis of speaker variation.

Limitations

• The dataset's geographic and demographic distribution within Turkey is unknown, which may introduce bias.
• Potential for label noise or inconsistency exists as the data is merged from 88 different sources.
• No information is provided on audio quality, length, or sampling rate uniformity.

Provenance

Source

Hugging Face dataset by Codyfederer, merging 88 source datasets.

Collection Method

Merged from 88 existing speech datasets; original sampling rate preserved.

Time Range

null

Freshness

Last updated on August 30, 2025.

Geography

null