Datasets & models

Datasets · 7

• ParsVoiceEMNLP 2026 Main

  A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis

  The largest Persian speech corpus built for TTS: audiobook speech aligned to text by an automated pipeline, with a BERT-based sentence-completion detector and binary-search boundary optimization. A fine-tuned XTTSv2 model reaches 3.6/5 naturalness MOS and 4.0/5 speaker-similarity MOS.

  2,178 hours of speech1.36M aligned segments1,803 speakers

  • Paper
  • Code
  • Dataset
  • Model
  • Demo

• PersianPuncSilkRoadNLP @ EACL 2026

  A Large-Scale Dataset for Persian Punctuation Restoration

  17 million Persian samples for punctuation restoration, released with a fine-tuned ParsBERT token-classification model (91.33% macro-F1) that restores punctuation without rewriting the text, unlike LLMs, which tend to over-correct.

  17M samples91.3 macro-F1

  • Paper
  • Dataset
  • Model

• PersianMedQALREC 2026

  A Persian-English Bilingual Medical QA Benchmark for LLMs

  20,785 expert-validated multiple-choice questions from 14 years of Iranian medical exams across 23 specialties, in Persian and English, with 41 LLMs evaluated.

  20,785 questions23 specialties

  • Paper
  • Dataset

• PARSA-BencharXiv preprint

  A Persian Audio-Language Model Benchmark

  The first benchmark for audio-language models on Persian language and culture: 16 tasks across speech understanding, paralinguistics and culturally grounded audio reasoning, including Persian poetry and music.

  16 tasks

  • Paper
  • Dataset

• PerCulNAACL 2025

  A Story-Driven Cultural Evaluation of LLMs in Persian

  Dataset designed to assess cultural sensitivity of LLMs toward Persian culture through story-based, multiple-choice questions.

  • Paper
  • Dataset

• PolitiSky24Findings of EMNLP 2025

  U.S. Political Bluesky Dataset with User Stance Labels

  First-of-its-kind dataset for stance detection on Bluesky, focused on the 2024 U.S. presidential election. Contains 16,044 user-target stance pairs with detailed metadata.

  • Paper
  • Dataset

• Persian Offensive Language Dataset2024

  Hate Speech and Offensive Language Detection

  Dataset of 38K Persian tweets containing hate and offensive language, manually annotated with bias measurement metrics.

  • Paper

Models · 3

• ParsVoice-XTTSEMNLP 2026 Main

  Multi-Speaker Persian Text-to-Speech

  XTTSv2 fine-tuned on ParsVoice for multi-speaker Persian speech synthesis with voice cloning.

  • Paper
  • Code
  • Model

• ParsBERT PunctuationSilkRoadNLP @ EACL 2026

  Persian Punctuation Restoration Model

  ParsBERT fine-tuned as a token classifier on PersianPunc to restore punctuation in raw Persian text.

  • Paper
  • Model

• ARMANEMNLP 2021

  Pre-trained Model for Persian Abstractive Summarization

  Transformer-based encoder-decoder model achieving state-of-the-art performance on Persian summarization tasks.

  • Paper
  • Code

Citing

If you use these resources, please cite the linked paper. BibTeX is on each paper's page and in the Hugging Face dataset cards.