Reading Hebrew with taamim tune — scoping notes
Prepared 2026-11-17, after the user picked this direction from the tiny-voca brainstorm (Nov 16 evening, "Reading Hebrew with taamim tune"). No building yet — direction first. This note answers the three open questions the assistant asked, with evidence, and recommends a first proof.
The core insight (why this is not just tiny-voca in Hebrew)
Trop is almost algorithmic. Each ta'am is a short melodic motif, and a pasuk's melody is the sequence of its motifs. So the tune may not need a neural model at all: a small voice for the syllables plus trop rules for the melody could be smaller than tiny-voca's 18 MB.
The hard part is the reverse of tiny-voca's trick. Tiny-voca worked because a bigger TTS existed to generate its 19 hours of training audio. No big TTS sings taamim — commercial Hebrew TTS reads niqqud but explicitly does not apply cantillation melody (SpeechGen: "Cantillation melody (ta'amei ha-mikra) is not applied"), and defaults to Modern Israeli phonology. So there is no teacher model to copy from. The melody has to come from somewhere else: either rendered from the marks by rule, or recorded.
Text: solved, open data
- Sefaria's default Tanakh text includes vowels and taamim, with toggles for each; its v3 texts API also offers a "Tanach with Text Only" version (no vowels, no taamim). Source: Sefaria help + developer portal.
- Miqra according to the Masorah (MAM), ed. Seth Avi Kadish, CC BY-SA 3.0 — the text Trope on the Go credits.
- tikkun.io Tikkun column dataset (akivajgordon/tikkun.io, MIT licence) — also credited by Trope on the Go.
- tanach.us is the other standard open Masoretic text source.
Text, nikud and taamim are not the bottleneck. Pronunciation is: Ashkenazi vs Sephardi/Modern Israeli is a real choice, and existing TTS defaults to Modern Israeli (final-syllable stress, merged gutturals).
Trop system: scope is bigger than "Torah vs Haftarah"
Eastern Ashkenazi tradition has six modes (Malin, Hebrew University): 1. Torah, normal Shabbat/weekday/festival 2. Torah, High Holidays (Rosh Hashanah / Yom Kippur mornings) 3. Haftarah 4. Lamentations (Tisha B'Av) 5. Esther (Purim) 6. Festive Megillot — Shir Hashirim, Rus, Koheles
Plus special melodies inside Torah mode for Aseret haDibrot and Az Yashir, and a coda motif for the last few words of each reading (and a different one at the end of each sefer).
Recommendation: scope v1 to Torah mode only, Ashkenazi. Haftarah is the natural v2 (same marks, different melodies). Megillot / Eicha / Esther are separate modes — defer, don't promise them in v1.
What already exists (the gap is specific, not empty)
Recorded-audio lein tools already exist and are good: - Trope on the Go (cantor-built, Ashkenazic Torah + Haftarah, 4,000+ word/phrase library, record-and-compare, animated sound maps) - Torah Morah (Sefaria-powered, recordings of every aliyah, iPhone) - PocketTorah, Trope Trainer, Virtual Tikkun Kahal (launched Nov 11, 2026, native apps, $10/mo)
None of them synthesises arbitrary text. Their audio is a fixed recording library. The gap a tiny-voca-style voice would fill: hear ANY pasuk or aliyah read correctly on demand — any aliyah, any week, including ones no recording library covers — offline, in-app, in a voice you own.
That is also the fit with what he already ships: Tikkun Korim (his Torah reading app — a Leonard Berkowitz print bug is open on it) is the natural home, with Shafeh's learning tools as the second home. A lein-practice tool (hear any aliyah, then try it yourself) is sharper than passive listening inside the Siddur.
Two architectures
A. Rule-rendered melody over a tiny syllable voice (recommended to test first). Record / synthesise one voice saying syllables, then pitch-shift and time-stretch each syllable along the motif for its ta'am. Needs a motif table: one short melody per ta'am, in Torah mode, Ashkenazi — a finite set, plus the verse-end coda and the special-section melodies. Risk: sounds chant-like but flat if the motifs are wrong; needs a cantor to verify the motif table by ear.
B. Neural TTS trained on sung audio. Needs a training corpus of sung taamim, which does not exist as open data. Would have to be recorded: a cantor singing the motif set, or full aliyot. That is a recording project before it is a model project. Do not start here.
Either way, the training-data question from the brainstorm stands: the synthetic recipe works for the syllables, not for the tune.
Recommended first proof (small, checkable)
One aliyah, Torah mode, Ashkenazi: Parshat Vayetzei, first aliyah (this week's parsha, Shabbat Nov 21, 2026). Steps: 1. Confirm the motif table for Torah mode (one melody per ta'am + coda) against a known cantor recording. 2. Render Vayetzei 28:10–12 (first few pesukim) with a placeholder voice + rule-rendered trop, and listen. 3. If it sounds like leining, the architecture works; if not, the motif table or the voice is the problem, and that is cheap to learn now.
Decisions still his (3)
- Torah only for v1, or Torah + Haftarah from the start? (Recommend: Torah only.)
- Which trop tradition? (Presumed Ashkenazi — his community, Chabad — but not confirmed. Sephardi is a different melody set.)
- What is it for — learn-to-lein (hear any aliyah, practise it) or listening inside Siddur/Tehillim? (Recommend: learn-to-lein, in Tikkun Korim.)
Also open: pronunciation for the syllable voice (Ashkenazi Hebrew, with its vowel values, vs Modern Israeli), and whether the voice should be a real cantor's recorded motifs or fully synthetic.
Sources checked 2026-11-17
- Sefaria help: Tanakh viewable with/without vowels and cantillation; developers.sefaria.org (v3 texts API, "Tanach with Text Only").
- Malin, "Eastern Ashkenazi Biblical Cantillation", Hebrew University (six modes listed above).
- Trope on the Go (tropeonthego.com): Ashkenazic Torah + Haftarah, 4,000+ word/phrase library, credits MAM text, Sefaria API, tikkun.io.
- Torah Morah (torahmorah.org): Sefaria-powered, recordings of every aliyah.
- SpeechGen Hebrew TTS page: niqqud read, "Cantillation melody (ta'amei ha-mikra) is not applied", Modern Israeli phonology default.
- Tiny-voca (Nov 16): 4.5M params, 18 MB, 19h synthetic training audio, ~$26, one RTX 4090 for a day; demo/model/code open.