New speech dataset aims to fill conversational gaps in audio AI models

Researchers have created MENASpeechBank, a new dataset designed to address a critical shortage of conversational speech data for training audio large language models (AudioLLMs). The collection focuses on understudied regional voices and dialects, particularly those from the Middle East and North Africa.

What is MENASpeechBank?#

The dataset comprises approximately 18,000 high-quality speech utterances collected from 124 speakers representing multiple countries across the MENA region. Coverage includes English, Modern Standard Arabic (MSA), and regional Arabic varieties—linguistic contexts where diverse training data has historically been scarce. The recordings serve as a foundation for generating larger synthetic datasets while maintaining authenticity.

How it works#

Rather than relying solely on crowd-sourced recordings, the researchers developed a controllable pipeline to expand the dataset systematically:

• Persona profiles are enriched using attributes inspired by the World Values Survey, grounding each speaker in cultural and value-based characteristics
• A taxonomy of ~5,000 conversational scenarios provides context for naturalistic multi-turn interactions
• Semantic matching pairs personas with scenarios to ensure coherent role-play conditions
• An LLM generates ~417,000 dialogue exchanges in which synthetic speakers take on persona-conditioned user roles while an AI assistant responds helpfully
• Speaker-conditioned audio synthesis creates realistic utterances for user turns while drawing from reference recordings, preserving individual vocal characteristics

Why it matters#

AudioLLMs have rapidly advanced in their ability to follow speech-based instructions, but training them remains hampered by the lack of abundant, diverse, instruction-aligned speech–text pairs. The problem is especially acute for persona-grounded and dialectal interactions, where collecting genuine multi-speaker conversations at scale remains expensive and time-consuming. MENASpeechBank bridges this gap by offering both synthetic and human-recorded dialogue options, along with detailed evaluation of their respective qualities.

The researchers indicate that MENASpeechBank will be made available to the research community, potentially accelerating progress in building audio AI systems that better represent underrepresented regions and languages.