ulaspolat/tsc-tr-filtered-94h

Dataset Card for Turkish Speech Corpus (TSC) — Preprocessed Edition

Dataset Summary

This dataset is a preprocessed and filtered version of the Turkish Speech Corpus (TSC), originally published by the Institute of Smart Systems and Artificial Intelligence (ISSAI) at Nazarbayev University. The original corpus contains 218.2 hours of transcribed Turkish speech across 186,171 utterances and is described in the paper Multilingual Speech Recognition for Turkic Languages (Mussakhojayeva et al., 2023).

This version applies a multi-stage quality filtering pipeline to the original audio, removes recordings that contain music or significant background noise, and augments each recording with a pseudo speaker label generated by unsupervised speaker clustering.

---

Original Dataset

Full credit for the underlying audio recordings and transcripts belongs to the creators of the Turkish Speech Corpus:

Mussakhojayeva, S., Dauletbek, K., Yeshpanov, R., & Varol, H. A. (2023). Multilingual Speech Recognition for Turkic Languages. Information, 14(2), 74. https://doi.org/10.3390/info14020074

The original dataset is available at: https://huggingface.co/datasets/issai/Turkish_Speech_Corpus

Dataset Structure

Data Splits

Split	Recordings
Train	72,245

Data Fields

Each example contains:

• audio — the audio recording (WAV, mono)
• text — the corresponding Turkish transcript (from the original TSC)
• speaker_id — a pseudo speaker label (e.g., speaker_42) assigned by unsupervised clustering; not a verified identity

---

Processing Pipeline

The pipeline applies several sequential stages to the raw TSC audio. Each stage is described below in the order it was applied.

Stage 1 — Music and Background Noise Filtering

The first stage removes recordings that contain music, significant background noise, or insufficient speech content. This is done using PANNs (Pretrained Audio Neural Networks), specifically the CNN14 model trained on the AudioSet ontology.

Each audio file is passed through the CNN14 classifier, which produces confidence scores across 527 AudioSet event categories. A recording is retained only if it satisfies all of the following:

• The Speech class score exceeds a minimum threshold.
• The Music class score falls below a maximum threshold.
• Background noise scores (covering a range of noise-related AudioSet categories) fall below their respective thresholds.
• The recording is at least a minimum duration.

Recordings that fail any of these checks are excluded from the dataset along with their paired transcript files.

Stage 2 — Speaker Embedding Extraction

For every recording that passes Stage 1, a fixed-dimensional speaker embedding is extracted using a deep speaker verification model. The pipeline was developed with pyannote/wespeaker-voxceleb-resnet34-LM — a ResNet34 model fine-tuned on VoxCeleb.

The model processes the entire audio file as a single segment and produces a single embedding vector per recording. Embeddings are extracted and saved for the clustering stage.

Stage 3 — Unsupervised Speaker Clustering

Speaker embeddings are L2-normalized and grouped using HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise), a density-based clustering algorithm well-suited to datasets where the number of clusters is unknown.

The clustering procedure:

1. All embeddings are clustered jointly.
2. Samples assigned to the HDBSCAN noise cluster (label -1) are discarded.
3. Clusters that contain fewer than a minimum number of recordings are further discarded as likely artefacts.
4. The remaining clusters receive sequential, anonymized labels: speaker_0, speaker_1, …, speaker_N.

The final speaker_id field in the dataset corresponds to these cluster identities. Because the labels are derived from unsupervised clustering, they represent acoustic speaker groups, not verified speaker identities. Two recordings with the same speaker_id are predicted to originate from the same speaker, but this prediction may occasionally be incorrect.

Stage 4 — Additional Quality Checks (Subset)

A secondary quality pass using pyannote.audio diarization models was applied to a subset of the data to evaluate:

• Single-speaker detection: Recordings estimated to contain more than one speaker were flagged.
• Overlapped speech detection (OSD): Recordings with simultaneous speech from multiple speakers were flagged.

These checks were conducted on a limited subset to audit pipeline quality rather than to filter the full dataset. As such, the final published dataset does not guarantee that all recordings are single-speaker or free of overlapping speech.

---

Speaker Labels

Speaker labels in this dataset are pseudo-labels produced by unsupervised acoustic clustering. They should be interpreted as follows:

• Same speaker_id → high likelihood of the same speaker.
• Different speaker_id → the model determined these recordings are acoustically distinct enough to belong to different speakers.

---

Limitations and Caveats

• Speaker labels are approximate. HDBSCAN clustering on speaker embeddings provides a good approximation of speaker identity but is not perfect. Cluster purity has not been formally evaluated on a held-out set.
• Removed recordings are not documented per-sample. The dataset does not include metadata indicating why a specific recording was excluded. Users needing the full original set should use the source dataset.
• No quality guarantee for individual recordings. Despite the multi-stage filtering pipeline, some utterances in the final dataset may still contain multiple speakers, background noise, or music. The PANNs-based filter operates at the file level using probabilistic confidence scores, and the diarization-based checks were applied only to a small subset of the data. No filter is perfect, and edge cases will inevitably pass through.
• Domain of speech. The speech content and recording conditions are those of the original TSC. This processing pipeline does not alter or augment the acoustic content.

---

Citations

If you use this dataset, please cite both this preprocessed version and the original paper:

Original Dataset and Paper

@Article{info14020074,
  AUTHOR  = {Mussakhojayeva, Saida and Dauletbek, Kaisar and Yeshpanov, Rustem and Varol, Huseyin Atakan},
  TITLE   = {Multilingual Speech Recognition for Turkic Languages},
  JOURNAL = {Information},
  VOLUME  = {14},
  YEAR    = {2023},
  NUMBER  = {2},
  ARTICLE-NUMBER = {74},
  URL     = {https://www.mdpi.com/2078-2489/14/2/74},
  ISSN    = {2078-2489}
}

---

Acknowledgements

This dataset was created by processing the Turkish Speech Corpus originally developed and published by the Institute of Smart Systems and Artificial Intelligence (ISSAI), Nazarbayev University.

The filtering and speaker-clustering pipeline uses the following open-source tools and pretrained models:

• PANNs (Pretrained Audio Neural Networks) — CNN14 model for audio tagging
• pyannote.audio — speaker embedding and diarization models
• HDBSCAN — density-based clustering for speaker grouping
• scikit-learn — supporting clustering utilities