Reza2kn/nasle-mana-clean

Nasl-e-Mana Speech Corpus

Clean, playable Persian speech audio collected from the public Nasl-e-Mana magazine WordPress site. The export contains two intentionally different collections:

Split	Rows	Columns	Meaning
labeled configuration (train/)	809	audio, label	Audio with recovered article text. These clips were identified as the consistent female source-text voice and are kept together.
to_transcribe configuration (to_transcribe/)	626	audio	Playable audio for which the source page did not provide a usable matching article text.

The audio column is a Hugging Face Audio feature, so clips should be playable directly in the dataset viewer and loadable with datasets.load_dataset. The two collections are exposed as separate configurations because the standard Hugging Face AudioFolder builder requires the same feature columns across splits; this preserves your requested exact schemas instead of adding an empty label column to to_transcribe.

Dataset creation

The source was crawled through the Nasl-e-Mana WordPress REST API and issue/article pages. Article text was recovered conservatively from visible post content while excluding image and media nodes. For paired clips, the label contains the linked article title, the displayed author line when present, and the visible body text in reading order.

Before this release, every exported file was checked with ffprobe. Of 1,512 discovered audio candidates, 1,435 decoded as playable MP3 and were exported. HTML responses, archive files, missing media, and other non-decodable objects were excluded. Audio files are preserved without transcoding; where the website supplied an MP3 with a misleading extension, the export exposes it with a normal .mp3 filename.

Important interpretation notes

• The train labels are alignment candidates, not verified ground-truth transcripts. They should be checked against decoded ASR before being used for a high-precision training set.
• The title and author are included because the recording commonly reads them before the article body. Users doing forced alignment or TTS cleaning may want to segment or remove those boundaries.
• to_transcribe is not a speaker-balanced evaluation split. Speaker identity and transcript quality for these clips remain unverified.
• The paired source-text clips are deliberately kept together in train because they share the same female speaker. Do not use this split as speaker-disjoint evaluation data.
• The crawl is a dated snapshot. The site may add, remove, or replace media after publication; the original source URL and provenance manifest were retained during collection.

Intended use

The dataset is intended for Persian ASR/TTS research, transcript verification, speech/audio quality analysis, and building a manually reviewed alignment corpus. Any public or commercial use should independently verify the rights to redistribute the source recordings and text.

Limitations and licensing

This dataset is derived from material published by Nasl-e-Mana. No separate redistribution license was identified during collection. The dataset card does not grant copyright permission; users are responsible for respecting the source site's terms, the rights of speakers and authors, and applicable law. Please contact the source publisher for permission where required.

Citation and source

Source website: naslemana.com

This release was prepared from a snapshot of the site's public content on 2026-08-29. The accompanying repository files contain the generated metadata and the audio files used for this export.