Unvocalized Persian sentences, built as the input pool for training a Persian diacritization (harakat) model. Text only — no audio. Total repo ~473 MB.

TL;DR — vocalizing? use `pool_100k`

from datasets import load_dataset
ds = load_dataset("MSVG1989/persian-text-pool-tts", split="train")   # default config = pool_100k
for r in ds:
    vocalize(r["id"], r["raw"])                # `raw` = the unmarked text

100,000 sentences. 50/50 formal-informal, 45% spoken / 55% audiobook, spread over all 11,012 books, mean 17.8 words. Nothing else needs to be filtered or balanced.

Repo layout

Every source lives in its own numbered folder, and every stage of the pipeline is kept, so any decision can be revisited without re-downloading 342 GB.

data/
  01_spoken_thomcles/          spontaneous spoken Persian (podcasts, interviews)
      full/          103613 rows   every extracted row, no filtering
      selected/       71786 rows   filters applied
  02_audiobook_fidibo/         audiobook prose, 11,012 books
      full/          312681 rows   raw audio-chunk rows, BEFORE sentence splitting
      sentences/     814773 rows   after splitting, BEFORE filtering
      selected/      517341 rows   filters applied
  03_pool_100k_balanced/       ** the deliverable **
      all/           100000 rows   the balanced pool
      from_spoken/    45000 rows   its spoken half, split out for inspection
      from_audiobook/ 55000 rows   its audiobook half
jsonl/     gzipped JSONL of every folder above, for line-oriented pipelines
code/      the entire pipeline that produced this
samples/   400 rows in plain text, if you just want to read some
MANIFEST.json  every count, filter, threshold, seed and source revision

Which config do I load?

config	rows	use it when
`pool_100k` (default)	100000	you are vocalizing. Balanced and ready
`pool_100k_from_spoken`	45000	you want to inspect or train on just the spoken half
`pool_100k_from_audiobook`	55000	same, audiobook half
`all_selected`	589127	you want the entire filtered pool, unbalanced
`spoken_selected`	71786	spoken Persian only
`audiobook_selected`	517341	audiobook prose only
`spoken_full`	103613	you want to change a spoken filter
`audiobook_sentences`	814773	you accept the sentence splitting but want different filters
`audiobook_full`	312681	you want to redo the sentence splitting itself

pool_100k* and every selected config share one identical schema, so they concatenate freely. audiobook_sentences and the two full configs carry their own source-shaped schemas.

full vs sentences vs selected vs pool

full — raw extraction, no filtering. Contains fragments, boilerplate and over-long rows. Kept so no filter decision is irreversible.
sentences (audiobook only) — audio chunks stitched and split into sentences, but before any filter. Kept because the splitting and the filtering are separate decisions and you may want to revisit only one.
selected — filters applied per source. Clean, but unbalanced: audiobook alone is 517341 rows, 84% formal, which would swamp the spoken register.
pool_100k_balanced — a balanced, diversity-spread draw from both selected sets. This is the deliverable.

Where it came from

Both sources were extracted from markmuller/fidibo_normalize_part1_part5, revision 91bebf08fc8628fb3f2e52a4ed67b7ff7bfb2edf.

That repo is 683 GB of 16 kHz TTS audio, not a text corpus. Text is roughly 0.07% of its bytes — 0.30 MB of text inside a 447 MB shard. Extraction meant streaming ~342 GB (measured 252 MB/s on 16 parallel connections, ~23 minutes), keeping the text columns and discarding the audio. It has no dataset card and no license; the only non-.arrow file at its root is .gitattributes.

Two upstream details worth recording:

Despite the name, parts 2, 3 and 4 do not exist in that repo — only 1 and 5.
Every corpus appears twice, as X_normalized and X_normalized_h. The _h variants carry text_harakat from an older, weaker model (confidence-filtered at min_conf=0.6) and were deliberately excluded — the whole point here is to re-vocalize from scratch. They also hold harakat_word_conf, an unfiltered per-word confidence array, which may be worth revisiting later for routing low-confidence words to an arbitrator model.

01 — `spoken_thomcles`

History, politics, economics and science podcasts and interviews: spontaneous spoken Persian, heavily colloquial (میره, ایشون, بخواد, می‌شه, رو) and largely unpunctuated, because it is transcribed speech. 390.9 hours of audio behind 103613 rows.

Rows kept whole, not split. With almost no punctuation, splitting yielded only 0.16 usable sentences per row; the median row is already 22 words.

02 — `audiobook_fidibo`

Audiobook prose: formal written narration with quoted colloquial dialogue, fully punctuated, 11012 distinct books.

These rows needed real work. The upstream rows are audio chunks, not sentences — median 46 words, up to 154, starting and ending mid-sentence. So per book we sort by audio segment index, stitch consecutive segments back together — never across an index gap, which means missing audio and would create a false join — split on .؟!؛, and keep only sentences with a real boundary at both ends. That turned 312681 chunks into 814773 candidate sentences at a median of 12–13 words.

The `text` field trap

Upstream ships three text fields and they are not interchangeable:

field	what it is
`text`	TTS-normalized: numbers verbalized (`۱۰` → `ده`), ZWNJ repaired (`خانه ها` → `خانه‌ها`). This is what we used.
`text_raw`	pre-normalization, keeps bare digits. Differs from `text` on 429/1062 spoken rows but only 5/485 audiobook rows
`text_plain`	punctuation stripped (audiobook only); differs on 485/485 rows

A diacritizer cannot mark digits, so text is the only correct choice. text_raw would have shipped numerals into the vocalizer; text_plain would have destroyed the punctuation that makes sentence splitting possible.

Fields

field	meaning
`id`	stable — `th-<shard>-r<row>` (spoken), `fid<part>-<book8>-s<n>` (audiobook)
`raw`	the sentence, harakat stripped — the vocalization input
`register`	`formal` / `informal` — a proxy, see limitation 2
`domain`	`spoken_podcast` or `audiobook`
`source`	upstream repo, config and revision
`n_words`	whitespace word count of `raw`
`selected_by`	full filter signature, for reproducibility
`register_score`	colloquial-marker rate the register label came from
`lex_coverage`	fraction of tokens in the project's 123,112-word lexicon
`had_source_harakat`	true if the upstream text already carried marks
`text_with_source_marks`	the original marked text, where it existed
`book`	book/recording id — null for every spoken row, see limitation 3
`duration_s`, `mos_ovr`	duration and DNSMOS of the aligned audio (spoken only)
`asr_cer_max`, `asr_wer_max`	worst ASR error among the chunks a sentence came from (audiobook only) — a text-reliability signal, not filtered on
`run_len`	how many audio chunks were stitched to make the sentence (audiobook only)
`src_shard`, `src_row`	exact provenance upstream

Selection

Seed 42 throughout; every row records its own selected_by.

step	spoken	audiobook
in	103613 rows	814773 sentences
length outside 8–30 words	−28463	−279822
lexicon coverage below floor	−2978 (floor 0.75)	−16651 (floor 0.80)
exact duplicate	−305	−760
near-duplicate (5-gram MinHash, J ≥ 0.55)	−81	−198
gold-494 exact / near	−0 / −0	−0 / −1
selected	71786 (69.3%)	517341 (63.5%)

The lexicon floors differ on purpose. The floor is set at ~P5 of each source's own coverage distribution — the same way the project's pipeline/collect.py calibrates per stratum. Spoken measured P05 0.765 → floor 0.75; audiobook sentences measured P05 0.818 → floor 0.80. Copying one number to both would have been wrong in one direction or the other.

The length filter is the dominant cut in both, deliberately: the long tail is audio-bounded runs that cannot be split reliably, the short tail is fragments like فصل نهم کتاب. Dedup is what removes podcast boilerplate — one sponsor read appeared 14 times, a show intro 11 times.

The held-out gold set is excluded

The 494 hand-reviewed sentences of the project's evaluation corpus are excluded by exact match and by token-set Jaccard ≥ 0.55. This removed 0 rows from spoken and 1 from audiobook — the gold set is news, forum, e-commerce and classical verse, which barely intersects podcasts and audiobooks. The check runs regardless, so the guarantee holds as sources are added. Never train on those 494; they are the ruler.

How the 100k was drawn

Quotas first, then round-robin across groups — one sentence per book/shard per pass — so the draw spreads as widely as the quota allows instead of emptying a few books:

cell	quota	drawn from	groups used	max per group
spoken / informal	30,000	54,581	97 / 97 shards	310
spoken / formal	15,000	17,205	97 / 97 shards	155
audiobook / informal	20,000	84,116	9,022 books	3
audiobook / formal	35,000	433,225	10,931 books	4

Result: 50,000 informal / 50,000 formal, 45,000 spoken / 55,000 audiobook, all 11,012 books represented, mean 17.8 words.

Spoken is held at 45% deliberately — it is the register the project's gold corpus has most of, and no earlier collection sampled real speech. The 17.8-word mean is close to the gold corpus's 17.3 on purpose: downstream per-sentence vocalization cost scales with length, and the cost model was calibrated on sentences that size.

Limitations — read before training on this

No literary or verse register exists here at all. The project's gold corpus has 108 literary sentences (poetry, lyrics, proverbs, classical verse). This pool is podcast speech and audiobook prose; nothing in it resembles them. Literary coverage must come from elsewhere — this pool cannot replace a Ganjoor-style collector.
The register labels are a proxy, not annotation. Colloquial-marker rate over tokens, thresholded at 0.02, validated against the 494 gold sentences: accuracy 0.756, informal precision 0.941, recall 0.629. So informal is reliable; formal is not — a short informal sentence carrying no marker lands in formal. The true informal share is higher than reported. Do not quote the register balance without this caveat.
Spoken rows have no per-show identifier. Thomcles_normalized ships neither metrics nor a non-null audio.path, so book is null for every spoken row and sentences cannot be capped per podcast. A prolific show may be over-represented invisibly. The only lever was a per-shard proxy across 97 shards, which is what the 100k draw uses. The audiobook source has no such problem — it has a real book key.
Sentence boundaries came from audio. Audiobook sentences are bounded at both ends by punctuation, so they are genuine sentences. Spoken rows are not — they begin and end where the audio segment did, which can be mid-sentence.
7.06% of spoken rows arrived already carrying harakat from the upstream pipeline. Stripped for raw, preserved in text_with_source_marks. Stripping also normalises a few letters, e.g. اصلاً → اصلا, کِیل → کیل.
ASR error metrics are carried but not filtered on. Audiobook text came from forced alignment against audio; asr_cer_max / asr_wer_max expose how bad the worst contributing chunk was (upstream median CER ≈ 0.023). If you want a stricter pool, filter on these — we deliberately did not, to keep the filter set to the ones we could justify from measurement.

Reproducing

code/ holds the whole pipeline:

python3 extract.py Thomcles_normalized extract/thomcles.jsonl   # stream, keep text, drop audio
python3 segment.py extract/fid1.jsonl  extract/fid1_seg.jsonl   # audiobook only
python3 select2.py extract/thomcles.jsonl spoken_podcast sel/thomcles --lex 0.75
python3 select2.py extract/fid1_seg.jsonl,extract/fid5_seg.jsonl audiobook sel/fidibo --lex 0.80
python3 pool.py    sel/pool_100k.jsonl
python3 publish3.py

Needs the project's words.jsonl (123,112 words) and corpus/corpus.jsonl (the 494 gold sentences) alongside, plus read access to the gated upstream repo.

License

The upstream repo carries no license and no dataset card. No license is asserted here and none is granted. This is derived text redistributed for research on Persian diacritization, gated pending clarification from the upstream owner. Ask before using it for anything else.

Downloads last month: -

Total file size:

473 MB

You need to agree to share your contact information to access this dataset

Persian Text Pool for TTS

TL;DR — vocalizing? use pool_100k