ashirbadsahu/banking-faq-hi-en-speech

Banking FAQ Hindi-English Speech Dataset

A multilingual banking FAQ dataset that combines an English FAQ source with Hindi translation and synthetic speech generation.

Overview

This project starts from a public English banking FAQ dataset and converts it into a Hindi-English speech dataset for research and experimentation in conversational AI and multilingual speech systems.

Source dataset

• —Kaggle: https://www.kaggle.com/datasets/rudrakumargupta/banking-faq-dataset-for-chatbot-training

Translation pipeline

The English FAQ data was translated into Hindi using Qwen/Qwen3.6-27B, executed locally with unsloth/Qwen3.6-27B-GGUF through llama.cpp.

Speech synthesis pipeline

The translated text was converted to speech using k2-fsa/OmniVoice via omnivoice.cpp.

Audio specification

• —Sample rate: 24 kHz

Data pipeline

1. 1.Collect English FAQ pairs from the Kaggle source.
2. 2.Translate the English questions and answers into Hindi using the local Qwen inference stack.
3. 3.Apply post-processing for consistency and quality checks.
4. 4.Synthesize spoken audio using OmniVoice.
5. 5.Package the bilingual text and generated speech into a usable research dataset.

License

This project is licensed under the MIT License. See LICENSE.md for the full text.

Limitations

• —The Hindi data is generated by machine translation and may contain translation artifacts or wording that is not perfectly natural.
• —The synthesized speech reflects the characteristics of the TTS model and may not match human speaker variability.
• —Banking terminology should be reviewed before use in production or high-stakes systems.
• —The dataset is synthetic and should be validated against real-world customer conversations before deployment.