KRASR/kazakh-russian-asr-dataset

Kazakh-Russian Mixed ASR Dataset

Dataset Description

This repository documents an internally curated dataset prepared for an academic thesis project on automatic speech recognition for Kazakh and Kazakh-Russian mixed speech.

The dataset was created to support experiments with multilingual and code-switching ASR models. It focuses on speech where Kazakh is the dominant language, while Russian words or phrases may appear naturally inside the utterance.

The main contribution of this dataset is the preparation workflow, including source selection, audio segmentation, transcription review, text normalization, split design, and evaluation setup for Kazakh-Russian mixed-speech recognition.

Dataset Composition

The dataset consists of internally curated Kazakh-Russian mixed-speech audio segments and corresponding normalized transcriptions.

The dataset was prepared specifically for ASR experiments and does not include speaker identity labels, demographic labels, or biometric annotations.

Purpose

The dataset was prepared for academic research on automatic speech recognition, with a focus on:

• Kazakh speech recognition
• Kazakh-Russian mixed-speech transcription
• Code-switching ASR evaluation

Access

The full dataset is not released as a fully open public corpus due to licensing and redistribution limitations of some source materials.

Access may be considered upon individual academic request for research, verification, or evaluation purposes only.

Users who request access should agree that the dataset will not be used for commercial redistribution, speaker identification, biometric profiling, or re-publication of raw audio materials.

Data Fields

The dataset manifests use the following fields:

• audio: path or reference to the audio segment
• text: normalized transcription
• source: language label mixed
• duration: segment duration in seconds

Example manifest entry:

{
  "audio": "data/train/audio_000001.wav",
  "text": "мен бүгін университетке бардым",
  "source": "mixed",
  "duration": 8.42,
}

Preprocessing

Audio and text were prepared for ASR experiments using a consistent preprocessing pipeline.

The preprocessing workflow included:

• audio segmentation into short utterances
• conversion to a standard ASR-compatible format
• transcription review and correction
• text lowercasing
• whitespace normalization
• punctuation cleanup
• preservation of Kazakh-specific letters
• train, validation, and test split preparation

The text normalization process was designed to reduce formatting noise while preserving the linguistic content needed for ASR evaluation.

Limitations

The dataset was prepared as an academic research dataset and has several limitations:

• the data may not represent all Kazakh-Russian speaking styles
• some utterances may contain informal or conversational speech
• transcription quality depends on the clarity of the original audio
• the dataset is not designed for speaker recognition or demographic analysis
• the dataset is intended for research and evaluation rather than production deployment

Ethical Considerations

The dataset was prepared for speech-to-text research only. Speaker identity, personal profiles, demographic labels, and biometric attributes were not used as training targets.

The dataset should not be used to identify speakers, infer personal characteristics, or build profiling systems.

Citation

If you use this repository, dataset description, or dataset preparation workflow, please cite the related academic thesis/project.

Contact

Access requests may be considered for academic research, verification, or evaluation purposes.