New AI Builds a Huge Free Voice Dataset for Persian Speakers

Building voice AI for low-resource languages is hard: large-scale, high-fidelity speech datasets are scarce, traditional alignment-based methods need rare verbatim transcripts, and standard in-the-wild pipelines often rely on single-model automatic speech recognition and silence-based segmentation — leading to transcription errors and truncated prosody. PersianVox addresses this for Persian with a fully automated pipeline that generates high-quality speech corpora from unlabeled web data, yielding a 2,400-hour multi-speaker dataset described as the largest open-source speech resource available for Persian to date. It combines a prosody-aware segmentation strategy that uses acoustic turn-detection to preserve linguistic completeness and optimize utterance duration for long-context modeling, with a dual-model agreement mechanism that leverages two distinct model architectures to filter unreliable transcriptions without ground truth. The paper also provides the first comparative benchmark of speech quality assessment methods for Persian, releasing a human-annotated subset to facilitate future research.