medresetuzzehra/indicative-readings-voice

Indicative Readings Voice (Harfî Okumalar Sesi)

540 machine-generated spoken clips (404.3 minutes) of texts from medresetuzzehra/indicative-readings version 3.10.1: 257 in Turkish, 188 in English, 29 in Spanish, 28 in Indonesian, 17 in Portuguese, 15 in German, 6 in French. Version 1.4.0, 2026-10-30.

Machine-generated audio. Every clip is synthetic speech produced by a text-to-speech engine. It is not a human voice and does not imitate any person. Please keep this notice when you share the clips.

What is here

File	Content
audio/<id>.wav	the clip: mono, 16-bit PCM, 44.1 kHz WAV, byte-identical to the file that passed the gate
metadata.jsonl	one row per clip (file_name is the audio path, so the folder loads as an audio dataset)
manifest.json	file list with SHA-256 hashes, selection rule, gate models and counts
LICENSE, README.md	licence and this card

Clips by record kind: frontier 26, law 56, profile 56, reading 402 (reading = a field reading, law = a law of nature and science, frontier = a frontier topic of today's science, profile = a field profile, spoken as two passages: the reality the field rests on and its answer to 'What is this universe?'). Engines and voices:

• KittenTTS, voice expr-voice-2-f: 2 clips
• Supertonic 3, voice F3: 1 clip
• Supertonic 3, voice M1: 537 clips

The metadata names, for every clip, the record in the text dataset (record_id, record_split), the spoken text (text, identical to the text of version 3.10.1), the engine, model revision and synthesis settings, the gate result and the SHA-256 of the audio.

How the clips were chosen

The project synthesised 2429 clips for this pilot (calibration trials included) and checked each one by speech recognition; 948 passed the gate's thresholds. Only the clips that pass the strictest criterion are published here (540 of 2429); the others are not published. A clip is included only if all of these hold:

1. Two recognisers (the deciding models) transcribe it with zero word edits against the spoken text: policy K053-v3 uses Systran/faster-whisper-medium and dropbox-dash/faster-whisper-large-v3-turbo. In English and Turkish, digits 0–999 are spelled out before comparing; the scripts of the other languages contain no digits, and no circumflex folding is applied outside English and Turkish.
2. The spoken text is identical to the text of the text dataset's version 3.10.1 (393 clips were synthesised from earlier script sets of versions 3.0.0 and 3.6.0 and 3.7.0 and 3.8.0; their text was compared with the current text and is identical).
3. The clip is labelled machine-generated.

Other recognisers are informational (informational_word_edits in the metadata is often above zero: the small and base models mishear words that both deciding models hear correctly, and the base model is weak in Turkish). Recognisers used, with revisions:

• Systran/faster-whisper-base @ ebe41f70d5b6
• Systran/faster-whisper-medium @ 08e178d48790
• Systran/faster-whisper-small @ 536b0662742c
• dropbox-dash/faster-whisper-large-v3-turbo @ 0a363e9161cb

What this does not tell you

The gate measures one thing: whether the words spoken are the words of the text. It does not measure naturalness, stress, how a specific name is really pronounced, or whether the meaning comes across when listened to, and no listening study is part of this release. Treat the clips as a multilingual machine-speech sample of the text dataset, not as a recitation reference.

Pronunciation of one term

The text datasets coin one term, harfî, which speech recognisers write in many ways. For 105 of these clips the project's pronunciation dictionary (K068) was in force: the engine was given the term in the spelling below, and no other word was changed. The published text and the text field keep the original spelling; the metadata fields pronunciation_dictionary, pronunciation_term and pronunciation_respelling name the dictionary, the term and the spelling that was spoken. The gate compared the recognised words with the original text, accepting for this term its own spelling without the circumflex and the spellings listed in the dictionary for the language (German hafi, French and Portuguese arfi, Russian харфи); this rule was fixed before the clips were made and is in the project's decision log (K068, K069). In the first scoring of the batch the rule was not applied to the French and Portuguese clips because of an error in the project's code; it was corrected from the unchanged recognition results before any clip was selected for this release (decision log K069). Spellings given to the engine:

• German (de): harrfi, 13 clips
• English (en): harfee, 25 clips
• Spanish (es): jarfi, 25 clips
• French (fr): the text as written, 6 clips
• Indonesian (id): harfi, 22 clips
• Portuguese (pt): the text as written, 14 clips

All other clips (those of earlier batches, the Turkish clips of this batch and clips whose text has no such term) were made from the text as written.

Use

• Keep the machine-generated notice; do not present the clips as human speech or as a real person's voice.
• The text dataset keeps a locked test split. 126 clips here belong to records whose record_split is test: keep them out of any training, exactly as the text. The project itself does not use these clips as training data.
• If a record is withdrawn from the text dataset, its clip is removed in the next version of this one.

Licence and engine terms

The project's contribution (selection, metadata, audio as produced) is under CC BY 4.0; the spoken text is under CC BY 4.0 of the text dataset. The engines keep their own terms: Supertonic 3 (Supertone/supertonic-3, model card) is released under the OpenRAIL-M licence, which among other things asks that machine-generated content be disclosed as such; KittenTTS (KittenML/kitten-tts-mini-0.8) is Apache-2.0. The engine licences do not by themselves settle every possible use of generated audio; read them if your use is unusual.

Version history

• 1.4.0 (2026-10-30): 147 more clips from a fifth batch (540 in total), all passing the same criterion (zero word edits in both deciding recognisers, text identical to the text dataset, machine-generated label). The batch is the first made with the project's pronunciation dictionary for the coined term 'harfî' (K068): the engine was given the term in a calibrated spelling or, in French and Portuguese, as written, and the gate accepted the spellings listed in the dictionary; the metadata of each such clip says so. It also adds the first clips in German, Spanish, French, Portuguese and Indonesian (15, 29, 6, 17 and 28 clips; Russian had none): 950 scripts were tried, 289 passed the gate's thresholds and 147 had no word edit. English and Turkish gained 26 clips each. The clips are tied to Indicative Readings 3.10.1, whose text of every record that has a clip is unchanged. Clips published in 1.0.0 to 1.3.0 are unchanged. The gate measures whether the words spoken are the words of the text, not naturalness or the real pronunciation of the term; no listening study is part of this release.
• 1.3.0 (2026-10-30): 120 more clips from a fourth capacity-planned batch, all passing the same criterion (zero word edits in both deciding recognisers, text identical to the text dataset, machine-generated label); 393 clips in total. The batch draws on the longer field readings (up to 140 words) and on the frontier readings added in Indicative Readings 3.8.0. The clips are now tied to Indicative Readings 3.8.0, whose text of every record that has a clip is unchanged. Clips published in 1.0.0, 1.1.0 and 1.2.0 are unchanged.
• 1.2.0 (2026-10-24): 93 more clips from a third capacity-planned batch, all passing the same criterion (zero word edits in both deciding recognisers, text identical to the text dataset, machine-generated label); 273 clips in total. The batch also covers frontier readings and laws, which are longer. The clips are now tied to Indicative Readings 3.7.0, whose text of every earlier record is unchanged. Clips published in 1.0.0 and 1.1.0 are unchanged.
• 1.1.0 (2026-10-24): 101 more clips from a second capacity-planned batch, all passing the same criterion (zero word edits in both deciding recognisers, text identical to the text dataset, machine-generated label); 180 clips in total. Clips published in 1.0.0 are unchanged.
• 1.0.0 (2026-10-24): First release, 79 clips.

Türkçe özet

medresetuzzehra/indicative-readings 3.10.1 metinlerinden 540 makine üretimi ses klibi (404.3 dakika): 257 Türkçe, 188 İngilizce, 29 İspanyolca, 28 Endonezce, 17 Portekizce, 15 Almanca, 6 Fransızca. Tüm sesler konuşma sentezi motorlarıyla üretilmiştir; insan sesi değildir, kimseyi taklit etmez, bu etiketi koruyarak paylaşın. Klipler yalnız şu ölçütü geçenlerdir: iki karar verici Whisper modeli sesi yazıya çevirdiğinde söylenen metinle sıfır kelime farkı, metin yayımlanmış sürümle birebir aynı. Kapı yalnız söylenen sözcüklerin metne uygunluğunu ölçer; doğallık, vurgu, özel adların gerçek telaffuzu ve anlam akışı ölçülmemiştir. Klip sayısı küçüktür: üretilen 2429 klipten 948 tanesi kapı eşiklerini geçti, yalnız 540 tanesi yayımlanır. Metindeki kilitli test bölümüne ait kayıtların klipleri (record_split = test) eğitime girmemelidir. Lisans: CC BY 4.0; motor lisansları ayrıca geçerlidir.