MenaSpeechBank

# MENASpeechBank **MENASpeechBank** is a MENA-centric speech resource for building and evaluating persona-conditioned, multi-turn conversational systems. It combines a reference voice bank, synthetic persona profiles, and large-scale conversation data to support AudioLLM research across Arabic dialects, Modern Standard Arabic (MSA), and English. ## At a glance - **17,641** reference utterances - **26.45 hours** of audio - **123** unique speakers - **6** language or dialect labels: `Arabic_Egyptian`, `Arabic_Gulf`, `Arabic_Levantine`, `Arabic_MSA`, `Arabic_North_Africa`, `English` - **469** persona profiles - **416,279** persona-grounded multi-turn conversations ## Development pipeline  The figure above summarizes the pipeline used to construct MENASpeechBank, from source speech selection and persona creation to scenario matching, conversation generation, and speech synthesis. ## Repository contents - [data/train.parquet](data/train.parquet): default Hugging Face split used by the data viewer, with playable audio previews and core metadata fields. - [data/metadata.csv](data/metadata.csv): flat metadata table for the reference voice bank. - [data/MENASPEECHBANK.json](data/MENASPEECHBANK.json): JSON version of the reference voice bank metadata. - [data/MENASpeechBank/](data/MENASpeechBank/): WAV files for the reference voice bank, sharded across subdirectories. - [personas.jsonl](personas.jsonl): synthetic persona profiles enriched with demographic, linguistic, value, and personality attributes. - [msa_conversations.jsonl](msa_conversations.jsonl): persona-grounded multi-turn conversations matched to scenarios and domains. - [menaspeechbank_xtts_colab_example.ipynb](menaspeechbank_xtts_colab_example.ipynb): Google Colab example showing how to use MENASpeechBank with XTTS. - [speechbank_pipeline.png](speechbank_pipeline.png): pipeline figure used in this dataset card. ## Quick start The easiest way to inspect the dataset is through the Hugging Face viewer, which exposes the `train` split with playable audio. To work with the metadata programmatically: ```python from huggingface_hub import hf_hub_download import pandas as pd metadata_path = hf_hub_download( repo_id="QCRI/MenaSpeechBank", repo_type="dataset", filename="data/metadata.csv", ) metadata = pd.read_csv(metadata_path) print(metadata.head()) ``` ## XTTS example notebook The notebook [menaspeechbank_xtts_colab_example.ipynb](menaspeechbank_xtts_colab_example.ipynb) provides a minimal Colab workflow for using MENASpeechBank with `xtts_v2`. It is designed for: - Google Colab with a GPU runtime - Python **3.11** - a small, reproducible demo rather than a full training pipeline The notebook walks through: - downloading `data/metadata.csv` - selecting one or two bilingual reference speakers - fetching a small set of reference clips from the dataset - generating sample MSA and English speech with XTTS - listening to the generated outputs directly in Colab ## Reference voice bank schema The main reference table in [data/metadata.csv](data/metadata.csv) contains: - `file_name`: relative path to the reference WAV file - `language`: language or dialect label - `segment_origin`: source dataset identifier - `speaker_id`: speaker identifier - `segment_id`: segment identifier - `segment_duration`: segment duration in seconds - `segment_path`: path to the reference WAV file The Hugging Face `train` split in [data/train.parquet](data/train.parquet) exposes the following fields: - `audio` - `language` - `segment_origin` - `speaker_id` - `segment_id` - `segment_duration` - `segment_path` ## Personas and conversations [personas.jsonl](personas.jsonl) contains persona profiles with speaker-linked demographic, language, value, and personality information, together with a first-person persona summary. [msa_conversations.jsonl](msa_conversations.jsonl) contains conversation records with: - `custom_id` - `persona` - `conversations` - `matched_domain` - `similarity` - `scoring` These conversations are matched to personas using semantic similarity between persona summaries and scenario descriptions. ## Intended use **Recommended uses:** - benchmarking AudioLLMs under multi-turn spoken interaction - studying robustness across dialects, accents, and speaking conditions - evaluating profile consistency and speech-conditioned personalization - building small reference-based synthesis or adaptation experiments **Not recommended:** - speaker identification or biometric profiling - impersonation or voice cloning of real individuals outside controlled research settings - deployment settings that could create privacy, consent, or safety risks ## Ethics and limitations - Persona profiles are synthetic and are not intended to represent real individuals. - Synthetic conversations and synthesized speech may differ from naturally occurring dialogue in prosody, spontaneity, and disfluency patterns. - Coverage can be expanded further across countries, dialects, speakers, and scenarios. - The dataset is intended for research use. ## Licensing This dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (**CC BY-NC-SA 4.0**). ## Citation If you use MENASpeechBank in your research, please cite: ```bibtex @article{ali2026menaspeechbank, title = {{MENASpeechBank}: A Reference Voice Bank with Persona-Conditioned Multi-Turn Conversations for AudioLLMs}, author = {Ali, Zien Sheikh and Bhatti, Hunzalah Hassan and Nandi, Rabindra Nath and Chowdhury, Shammur Absar and Alam, Firoj}, year = {2026}, note = {arXiv:2602.07036 [cs.SD]}, url = {https://arxiv.org/abs/2602.07036}, } ```