Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·21d ago·5 min read
MENASpeechBank: A Reference Voice Bank with Persona-Conditioned Multi-Turn Conversations for AudioLLMs
TL;DR
MENASpeechBank is a new reference voice bank featuring approximately 18,000 high-quality utterances from 124 speakers across MENA countries, designed for training audio large language models (AudioLLMs). It enables diverse, persona-conditioned multi-turn conversations, addressing the scarcity of conversational speech data in various dialects.
✦ Why It Matters
Explore MENASpeechBank to enhance your audio model training with diverse, persona-driven conversational data.
Key Takeaways
- Utilize MENASpeechBank for training audio models with diverse speech data.
- Incorporate persona profiles to enhance user interaction in applications.
- Leverage the 5,000 defined conversational scenarios for varied training.
- Evaluate synthetic audio against human recordings for quality assurance.
How It Works
MENASpeechBank employs a controllable data pipeline that constructs persona profiles enriched with attributes from the World Values Survey. It defines a taxonomy of conversational scenarios and matches personas to these scenarios using semantic similarity.
The system generates role-play conversations where users adopt specific personas, and the assistant responds accordingly. This approach allows for the creation of synthetic audio that maintains speaker diversity, crucial for training robust AudioLLMs.