uk-en-code-mixed-asr-2h

A 2-hour dataset of Ukrainian-English code-mixed speech for automatic speech recognition, recorded by a single male speaker across 16 speaking-style personas covering software engineering domains (.NET, React, DevOps, project management).

• Samples: 448
• Total duration: 2h 6m 35s
• Language: 446 mixed (code-mixed) + 2 uk (monolingual)
• Speaker: 1 male voice, 16 personas (varied pacing, formality, topic; 15 map to Valera, 1 to ValeraAlt)
• Audio format: OGG Opus, 48 kHz mono
• EN level: C1 (all speakers)
• CS type: 446 intra-sentential, 2 monolingual
• Metadata: JSONL
• Pronunciations: Optional companion transliteration map (JSON)

Dataset Structure

.
├── README.md
├── metadata.jsonl
├── transliteration_map.json      (optional — see below)
└── audio/
    ├── 1_azure.ogg
    ├── 1_dotnet_base.ogg
    └── ...

metadata.jsonl

Field	Type	Description
id	string	Unique utterance ID (e.g. 1_azure)
file_name	string	Relative path to audio file (HF compatible)
transcription	string	Text with original code-mixing
duration	float	Duration in seconds
language	string	mixed (code-mixed) or uk (monolingual)
cs_type	string	intra (intra-sentential) or mono (monolingual)
speaker_id	string	Anonymized speaker ID (Valera / ValeraAlt)
en_level	string	English proficiency (C1)
topic	string	Text category / domain
translation_en	string	Full English translation (optional)

transliteration_map.json

An optional companion file mapping English terms found in the transcriptions to common Ukrainian phonetic transliterations. Useful for evaluating ASR models on pronunciation variants of code-mixed speech.

{
  "azure": ["ежур"],
  "function": ["фанкшн", "фанкшен"],
  "entity framework": ["ентіті фреймворк", "ентити фреймворк"],
  ...
}

Each key is an English term (lowercase), each value is an array of 1+ Ukrainian pronunciation variants sorted by commonness. The map covers 1,241 terms extracted from the transcriptions. It is intentionally kept as a separate file so it remains optional:

AI-generated content. This map was produced with the assistance of LLMs and may contain inaccuracies, inconsistencies, or unnatural transliterations. Use carefully.

• Without the map: use the transcription field as-is (original code-mixed text).
• With the map: dynamically generate alternative full-Ukrainian transcriptions by substituting English terms with their pronunciation variants, then evaluate ASR output against each combination.

Personas

Persona	Samples	Topic
dotnet_short	106	Short .NET standup updates
dotnet_react	101	.NET, React (daily updates, debug sessions)
pm	20	PM: delivery, planning, grooming, roadmap, retrospectives
devops	20	DevOps: CI/CD, Kubernetes, Terraform, AWS, observability
dotnet	20	.NET development
qa	20	QA: Playwright, load testing, security testing, test design
python	20	Python: scripting, async, FastAPI, pandas, data pipelines
java	20	Java: Spring Boot, microservices, JVM tuning, Gradle
cpp	20	C++: memory management, RAII, STL, CMake, performance profiling
sql	20	PostgreSQL/SQL: indexes, query optimization, migrations, transactions
ml	20	ML/MLOps: model training, MLflow, feature engineering, deployment
react_short	20	Short React standup updates
azure	19	Short Azure standup updates
devops_base	10	DevOps
react_dotnet	6	React, .NET replies
dotnet_base	6	.NET, Database standup updates

Usage

from datasets import load_dataset

dataset = load_dataset("vnikitin/uk-en-code-mixed-asr-2h", split="train")

Citation

If you use this dataset, please cite it as:

@dataset{nikitin2026uk-en-code-mixed-asr,
  author       = {Ruslan Valerii Nikitin},
  title        = {{uk-en-code-mixed-asr-2h}},
  year         = {2026},
  publisher    = {Hugging Face},
  version      = {1.0},
  url          = {https://huggingface.co/datasets/vnikitin/uk-en-code-mixed-asr-2h},
}

License

This dataset is provided under a custom ASR-only license.

Permitted use

• Automatic Speech Recognition (ASR) training, evaluation, benchmarking, and research
• Commercial and non-commercial ASR systems
• Linguistic analysis related to ASR, transcription, pronunciation, or Ukrainian-English code-mixed speech

Prohibited use

• Text-to-Speech (TTS) training
• Voice cloning, speaker imitation, voice conversion
• Speaker identification or speaker verification
• Biometric identification
• Generating synthetic speech that imitates or resembles the speaker
• Any non-ASR purpose without explicit written permission from nikitinrus404@gmail.com