jayant-havare-ibm/multi-ling-code-speech

Dataset Summary

This dataset contains 500 human-recorded, multilingual speech queries related to programming and code. It was created to evaluate the robustness of speech recognition and code-aware transcription systems under realistic human-speech conditions.

Unlike fully synthetic speech datasets generated using text-to-speech (TTS) systems, this dataset consists of naturally spoken queries recorded by human participants. Human speech introduces greater variability in pronunciation, speaking style, pauses, accents, and other characteristics that are typically underrepresented in synthetic speech.

The dataset covers five languages: English, Hindi, Marathi, Gujarati, Bengali

Data Collection:

Five participants were recruited for the data collection process. Each participant was fluent in one of the five target languages. Each participant recorded 100 spoken queries related to programming and coding in their respective language:

English: 100 samples Hindi: 100 samples Marathi: 100 samples Gujarati: 100 samples Bengali: 100 samples

This resulted in a total of 500 human-recorded audio samples.

Intended Use:

This dataset is intended for:

Evaluating multilingual speech recognition systems. Evaluating speech-to-code and code-aware speech transcription systems. Measuring ASR robustness on human speech. Studying multilingual programming-related speech recognition. Evaluating transcription refinement techniques. Comparing ASR systems on code-related queries. Research into speech interfaces for programming and developer tools.

Citation:

If you use this dataset in your research, please cite the associated work:

@dataset{human_speech_code_queries, title = {multi-ling-code-speech}, author = {Jayant Havare, Srikanth Tamilselvam, Ashish Mittal}, year = {2026}, publisher = {IBM, India.}, note = {Multilingual human-recorded speech dataset} }

Acknowledgements:

We thank all participants who contributed their speech recordings to this dataset.