Datasetpublic

Reza2kn/ganjoor-recitations

Ganjoor Persian Poetry Recitations (Full) Every published audio recitation on Ganjoor / AVA paired with its transcription — 30,133 clips, 1,276 hours of audio. Audio is stored full-length and unchunked, and every clip carries a single clean transcription in text, so it's ready for ASR / TTS training as-is. Columns column description audio full-length mp3 (native sample rate), embedded and playable text full transcription of the clip… See the full description on the dataset page: https://huggingface.co/datasets/Reza2kn/ganjoor-recitations.

sourceHugging Facecc-by-3.0updated 3mo agoView on Hugging Face

3likes506downloads

Dataset Card

Ganjoor Persian Poetry Recitations (Full)

Every published audio recitation on Ganjoor / AVA paired with its transcription — 30,133 clips, 1,276 hours of audio.

Audio is stored full-length and unchunked, and every clip carries a single clean transcription in text, so it's ready for ASR / TTS training as-is.

Columns

column	description
`audio`	full-length mp3 (native sample rate), embedded and playable
`text`	full transcription of the clip

What's in `text`

text is the poem body (from Ganjoor's authoritative written text, so archaic spelling is correct). Where the reciter speaks a preamble before the poem — a title such as حکایت سیمرغ, or an announcement such as غزل شمارهٔ ... دیوان شمس حضرت مولانا ... — that preamble is prepended as the first line, so the transcription matches what is actually said. The preamble is transcribed by Shenava-Koochik v1.0 (first 16 s of audio); ~13.2 k clips have one, the rest open straight into the poem or on music/silence.

Usage

python

from datasets import load_dataset
ds = load_dataset("Reza2kn/ganjoor-recitations", split="train")
ds[0]["audio"]   # {'array': ..., 'sampling_rate': ...}
ds[0]["text"]    # spoken preamble (if any) + poem

Notes

—30,133 recitations sourced from the Ganjoor public API (/api/audio/published). Full-length audio, no chunking. Of 30,136 published recitations, 3 are excluded because their source files are broken on the server (2 empty, 1 non-mp3).
—Some older recitations were re-encoded by Ganjoor after their API checksums were recorded; every file here was re-verified as a valid, complete mp3.
—The spoken preamble is machine-transcribed (Koochik v1.0). It is only added when the poem's opening was located after it (so it isn't a duplicate of the poem), but it can contain ASR spelling errors. The poem body itself is never ASR — it is Ganjoor's written text.
—Provenance for the merge is in the repo: manifest/manifest.jsonl (source metadata), intro/intro_all.jsonl (raw + ITN first-16s transcripts), merge/preambles.jsonl (exactly what was prepended, by id).

License & attribution