Learning to Lein Torah in Chabad Trop
Research findings for a lein-practice tool: Torah only, Chabad trop, pick a parsha and aliyah, hear it leined correctly, follow along, repeat. Done 17 November 2026. No building yet — this page publishes what was learned, including what is still uncertain.
Bottom line
- The grammar is fully documented; the melody is not. The taamim system — which mark, what it means, how phrases parse — is written down precisely enough to implement from text. The Chabad melody itself exists only in recordings (R' Michoel Slavin, baal korei of 770, on Chabad.org) and inside TropeTrainer's "Chabad" voicing system. There is no published, note-for-note "Chabad trop" chart.
- The text data is solved. Sefaria's API serves the whole Chumash with taamim (public domain), Hebcal has aliyah divisions as open data, and the Unicode encoding rules are known — including the traps (never NFC-normalize; the zarqa/tsinnorit swap).
- No TTS in the world sings taamim today. But a tiny-voca-style model can be driven to sing: it accepts per-phoneme duration and per-frame pitch overrides and sings within ~2 cents of the notes. A synthetic Chabad voice is plausible — it needs a validated Chabad motif table and an Ashkenazi-pronunciation Hebrew voice, neither of which exists in open form today. A recorded baal koreh is proven and authentic from day one.
- The gap is a Chabad-first teacher, not a player. Chabad.org's trainer is free and authentic but only plays recordings. TropeTrainer is the best generalist (Chabad is 1 of 79 systems). Nobody teaches the Chabad system clause-by-clause with phrase looping, lookalike drills, record-and-compare with pitch feedback, and a tikkun-to-STaM progression.
Motif table validation — what was measured
The working motif table (23 marks, transcribed from an Eastern Ashkenazi sheet-music chart) was checked against Slavin's recordings — 8 parshiyot, pitch contours extracted and cross-checked with three estimators. Results:
- 12 of 23 motifs validated as-is (shape matches): etnachta, zakef katan, tipcha, revia, tevir, pashta, munach, mahpach, merkha, darga, geresh, sof k'ria.
- 1 corrected: sof pasuk. The chart ends the verse on the tonic; Slavin's Gen 1:1 ends on a sustained low note (A2, a fifth below the operational tonic). The 2-2-3 rise is there, but the cadence lands low.
- 10 still uncertain — no clean isolated measurement: zakef gadol, yetiv, segol (its B♭ unconfirmed), zarka, kadma, telisha gedola, telisha ketana, gershayim, kadma-v'azla, pazer. They occur in the audio but run together with neighbouring phrases (Slavin pauses at only ~70% of phrase boundaries).
- Rare taamim, first measurements: shalshelet — "basically 2 pazers" is right in broad shape (two rising melismatic units) but wrong at note level; karnei parah and yerach ben yomo form a compound melisma with two high plateaus; mercha kefula (Lev 10:1) is a high melisma peaking at degree 6. One mercha kefula measurement (Gen 27:25) contradicts the other and is treated as unreliable.
The tonic is not pinned down — read this before using the table. Two readings of the tonic both fit part of the audio and neither fits all of it: D3 explains the reciting band and phrase endings; A2 explains the verse- and parsha-final cadences. Degrees in the table are expressed relative to D3 by convention. If the true tonic is A2, every degree shifts and the sof-pasuk "correction" disappears. On top of that, Slavin's intonation is approximate, not equal-tempered — phrase-final pitches form a continuum, not discrete semitones. What was validated is melodic shape (ups, downs, leaps, cadences), not exact pitches. A cantor's confirmation is the only thing that settles the degrees.
Recommended path
- Validate the motif table (done here, in part — see the caveats above).
- Build the shared infrastructure: text pipeline, aliyot, phrase parser, practice loop. No voice decision needed for this.
- Ship with a recorded voice — fast, authentic, zero ML risk.
- Prototype the synthetic voice in parallel; ship it only when a cantor validates its melody. A slightly-wrong synthetic chant teaches wrong melody, which is worse than a recording for a learner.
First proof proposed: one aliyah, Torah mode — Parshat Vayetzei, first aliyah (Gen 28:10–12).
The findings
- Research report — the taamim system from first principles, the Chabad melody, text data, existing tools, the voice fork, and the recommended path.
- Validated motif table — the 23 motifs with measured values from Slavin's recordings, verdicts, what changed, and what remains uncertain.
- Scoping notes — why the tune may not need a neural model, Torah mode v1, the two architectures, and the Vayetzei first-aliyah proof.
- Chabad trop melody deep dive — the full write-up behind the motif transcription, with sources.
- Source files — the markdown reports and the validation scripts and data (parser, pitch measurement, segmentation, whisper alignments, CREPE tracks). The audio recordings used for validation are Chabad.org's and are not included.