MODEL_CARD.md

metadata

license: mit
library_name: other
base_model:
  - Qwen/Qwen3.6-27B
  - k2-fsa/OmniVoice
pipeline_tag: text-to-speech
tags:
  - banking
  - faq
  - multilingual
  - hindi
  - english
  - speech
  - text-to-speech
  - synthetic-data
  - conversational-ai
model_name: Banking FAQ Hindi-English Speech Dataset

Model Card: Banking FAQ Hindi-English Speech Dataset

Model summary

This dataset is a multilingual banking FAQ corpus created by translating English FAQ content into Hindi and generating spoken audio for the translated text. It is designed to support research in multilingual conversational AI, synthetic speech generation, and domain-specific customer support systems.

Description

The source corpus comes from a Kaggle banking FAQ dataset for chatbot training:

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

The original text was in English. To create a Hindi-language version, the dataset was translated using Qwen/Qwen3.6-27B, run locally with unsloth/Qwen3.6-27B-GGUF via llama.cpp.

After translation, the Hindi text was converted to audio using k2-fsa/OmniVoice through omnivoice.cpp. The generated speech has a sample rate of 24 kHz.

Intended use

This dataset is intended for:

• research in multilingual banking QA
• customer support and FAQ experimentation
• TTS and speech synthesis evaluation
• ASR training on domain-specific spoken banking content
• conversational AI prototyping for Hindi-English systems

Out-of-scope use

This dataset should not be used for:

• deceptive voice cloning or impersonation
• harmful automated customer interactions
• production customer support without human review
• safety-critical financial decisions without oversight

Dataset structure

The dataset contains banking-related FAQ text and associated synthetic speech generated from the translated Hindi content. The primary language pair is English and Hindi, with audio generated in the spoken Hindi domain.

Training and generation pipeline

1. Collect English FAQ data from the Kaggle source dataset.
2. Translate questions and answers into Hindi using Qwen/Qwen3.6-27B.
3. Run the model locally using unsloth/Qwen3.6-27B-GGUF and llama.cpp.
4. Convert the translated text into speech using k2-fsa/OmniVoice via omnivoice.cpp.
5. Store the generated audio at 24 kHz sample rate.

Limitations

• Machine translation may introduce wording errors, unnatural phrasing, or domain mismatches.
• Synthetic speech may have pronunciation, prosody, or expressiveness limitations compared to human speech.
• The model may not reflect real-world speaker variation or emotional nuance.
• Banking terminology should be manually reviewed before using the data in sensitive or production workflows.

Bias, risks, and ethical considerations

This dataset is synthetic and domain-specific. It may reflect the translation model's stylistic and lexical preferences and may not generalize equally to all banking scenarios or dialects. Use responsibly and validate results on target populations and use cases before deployment.

Licensing

This project is released under the MIT License. See LICENSE.md for details.

The source Kaggle dataset and the underlying model dependencies may have their own licenses and usage terms. Please review those terms separately before redistribution or commercial use.

Citation

If you use this dataset, please cite the original source dataset and the model dependencies used in its creation.

Suggested attribution:

• Source dataset: Banking FAQ dataset for chatbot training (Kaggle)
• Translation model: Qwen/Qwen3.6-27B
• Local inference backend: unsloth/Qwen3.6-27B-GGUF with llama.cpp
• Speech model: k2-fsa/OmniVoice
• Speech runtime: omnivoice.cpp

Additional notes

This project is intended as a research and prototyping resource for Hindi-English banking speech generation and multilingual customer service workflows.