bcv-commons/hebrew-speaker-layer

Hebrew Bible Speaker Layer

Who speaks where in the Hebrew Bible: every quotation, its speaker, its edges on the Hebrew words, and the quotations inside quotations, on MACULA Hebrew's words and on the clause segments of the Hebrew Bible Clause Layout (which carries these columns on each segment since v0.37.1). bcv-commons/passage-facts builds on it: the verb that introduces each speech, the narrator's voice, the force of each clause, song and poetic lines, each tied to its words and source.

Release	v0.15, built 2026-10-27
Quotations	3,892 outer, 1,278 inside other quotations
Clause segments in a quotation	46,195 of 89,940
Clause segments from	the clause layout's segmentation 8977039d6ddb5383 (the same in every clause-layout release since v0.22)
Review	345 reviewed changes, each with its reason (review.tsv)
Speakers named in Hebrew	4,807 of 5,170 of the quotations, from the text itself
Characters	855 people and groups (speaker_id) over 4,551 quotations; 229 of them with a proper name (speaker_name, on 3,284 quotations); 87 groups where several speak, on 116 quotations (48 with speaker_members)
Changes	v0.15: speaker_id agrees with the quotation's own inquit (review rows where the speaker data put a line in the wrong mouth, or the inquit named the wrong person); where several speak, a group id and its members (speaker_members). v0.14: who the speaker is across quotations, as keys: speaker_id (one person, one id) and speaker_name / speaker_name_ref (that person's proper name in the text, רָחָב for the woman of Josh 2); speaker unchanged. v0.13: a quotation with no inquit of its own takes its speaker's Hebrew name from another quotation of the same speaker (speaker_basis); alt_eng keeps the English label only where no Hebrew name is found. v0.12: speakers and frames named in Hebrew from the text (speaker, speaker_ref, frame); alt_speaker and speaker_hebrew gone
Licence	CC BY 4.0

Files

• quotations.tsv: one row per quotation.
• segments/<BOOK>.tsv: one row per clause segment of the clause layout.
• corrections.tsv: the hand corrections, each with the reason the Hebrew text gives.
• review.tsv: the changes from the review of doubtful passages, each with its reason, its source and a confidence.
• prophets.tsv: the prophet of each prophetic book and the word of its heading that names him (the frame there).
• parallels.tsv: where parallel texts (Samuel-Kings and Chronicles, Isa 36-39 and 2 Kgs 18-20, Ps 18 and 2 Sam 22 …) differ in level or speaker.
• report.md: the build's figures and its lists of doubtful places.
• hebrew-speaker-layer-v0.15.tar.gz: all of the above in one pinned download; SHA256SUMS.

quotations.tsv

Column
id	the quotation: q… from Clear's data, L… added from the clause layout, S… a song or speech, N… a song its inquit announces, W… a quotation the layer puts around others (God's words in the prophets, a relayed message, a psalm opened by an inquit), R… added by a review row; a suffix (.1, .m2, .t3, .i4, .r5) marks a quotation inside the one before it
level, parent	1 for an outermost quotation; a quotation inside another is one level deeper and names it as parent
speaker	who speaks, named from the Hebrew text: the subject of the verb that introduces the quotation, as MACULA lemmas (יהוה, מֹשֶׁה, דָּוִד, מֶלֶךְ, אִשָּׁה), with the nouns of its construct chain (מַלְאָךְ יהוה, בֵּן יִשְׂרָאֵל, מֶלֶךְ אַשּׁוּר); for "the word of X" or "the oracle of X" (וַיְהִי דְבַר־יְהוָה … לֵאמֹר, נְאֻם יְהוָה) X; several subjects separated by commas. Where the quotation's own inquit names no one, the speaker's usual Hebrew name from another quotation (see speaker_basis); else empty
speaker_ref	the word keys of the words that name the speaker (in another quotation where speaker_basis says so)
speaker_basis	where speaker comes from: inquit (the quotation's own inquit), review (a review row), same speaker in book / same speaker in Bible (another quotation of the same speaker, see below); empty where there is no name
alt_eng	the speaker data's English label, only where speaker is empty and the label names someone (251 quotations: "shepherds at well", "watchman", "daughter of Zion"), so that no name the earlier releases gave is lost; empty everywhere else
speaker_id	who the speaker is across quotations: a stable id (c0124) for one person or group, the same in every quotation they speak and in later releases. Before v0.12 the label did this (one person, one label across a story); speaker names each quotation by its own inquit, so one person can be אִישׁ in one and עֶבֶד in the next, and one word stands for many people. A person with a proper name has one id across the Bible, one without (an unnamed servant, "the people") one per book. Empty for a speaker the speaker data names no one by ("he", "they", the one addressed, an untitled psalm)
speaker_name	that person's proper name as the text gives it (MACULA lemma; God's name for God): רָחָב, אַבְרָהָם, יהוה; empty for a person the text gives no name (Abraham's servant in Gen 24)
speaker_name_ref	the word key of a word that gives that name: the quotation's own inquit where it does, else the nearest quotation's that does (in the same book first), else the word a hand table gives (Josh 2:1, וּשְׁמָהּ רָחָב)
speaker_members	where several speak together (Rachel and Leah, Gen 31:14; Moses and Aaron): speaker_id is the group's and this lists the ids of the people its inquit names who are characters of their own and speak in that book, space-separated; the group itself has no speaker_name. Empty for one speaker
divine	Y when the speaker is God (decided in the build, also where speaker is empty)
quote_type, basis	where the quotation comes from: Clear's data (block, referent, character), the clause layout (layout), the translations' votes (translations), a saying verb inside a quotation (inner inquit), God's words in the prophets (oracle), a messenger formula (messenger formula, relayed message, commanded message), a song (song, voice), a psalm title (titled psalm), a hand correction (correction), a review row (review)
first, last	the first and last word, MACULA word keys (12 digits: book, chapter, verse, word, part)
from, to	the verses, Hebrew numbering
start_fit, end_fit	how the edges fit the clause segments
inquit	the line that opens a nested quotation (its verb also shows who speaks where speaker is empty: וְאָמַרְתָּ, וַיֹּאמְרוּ)

segments/.tsv

Column
chapter, verse, seg, clause_key	the clause segment, as in the clause layout
quotation, level, speaker	the innermost quotation the segment stands in, its level (0: none), its speaker
share	the share of the segment's words in that quotation
frame	the speaker of a text that is speech from its first line, as a field rather than a level, by the Hebrew lemma of the word that names them: a psalm's psalmist (the name in its title, לְדָוִד: דָּוִד; empty for a psalm without one); in the prophetic books the prophet, as the book's heading names him (יְשַׁעְיָהוּ, יִרְמְיָהוּ); Nebuchadnezzar's letter (Dan 4, נְבוּכַדְנֶצַּר of 3:31) and Daniel's visions (Dan 8-10, דָּנִיֵּאל of 8:1); a first-person account's narrator (Ezra 9, Neh 13): אֲנִי

Every segment's quotation is in quotations.tsv at the segment's level, and every parent holds its quotation.

Two readings of the prophets

In a prophetic book the frame is the prophet, and God's words are always a quotation (level 1, or deeper inside a message). So the levels can be read two ways:

• levels alone: the book as a text with God's words quoted in it (Ezek 6:2 level 1, Amos 3:8, the prophet's own words, level 0);
• the frame counted as a level: the book as the prophet's speech with God's words quoted in it, as unfoldingWord's translationNotes read it (Ezek 6:2 level 2, Amos 3:8 level 1).

Translations and reference works differ here, from book to book and within a book; the layer carries both and leaves the choice to the user.

How it is made

1. The outer quotations are Clear's "MACULA Quotation and Speaker Data": quotation spans with speakers, projected from an English translation's quotation marks onto MACULA's Hebrew words through the word alignment, with one speaker per quotation.
2. Refined on the Hebrew:
   • a quotation that starts on its own inquit ("וַתֹּאמֶר", "לֵאמֹר") starts after it, and one that ends on the next inquit ends before it;
   • an edge that only leaves out a prefix, particle or suffix of a clause segment ("הֲ", "וְ", the article, "כִּי", "הִנֵּה", a pronoun suffix: the English word order put the quotation mark there) moves to the segment's edge;
   • pieces of one speech are joined (the same speaker, nothing between them but a speaking formula);
   • quotations Clear lacks are added where the clause layout has speech with a first- or second-person verb (not in the first-person memoirs of Ezra and Nehemiah), and a run of three or more verses in the first or second person outside any quotation is a song or speech (1 Chr 16:8–35); a song its inquit announces ("אֶת־דִּבְרֵי הַשִּׁירָה הַזֹּאת", Deut 31:30) is one quotation of its singer.
3. Quotations inside quotations, from three kinds of evidence:
   • the Hebrew structure: a speech the clause layout opens inside a speech when the line it hangs on introduces speech (a verb of saying, "לֵאמֹר", a messenger formula); a saying verb inside a quotation followed by direct speech ("וָאֹמַר", Deut 32:20); a messenger formula "כֹּה אָמַר X" inside a quotation; the relayed message, where "כֹּה אָמַר X" comes before a quotation (Josh 24:2, Judg 11:15: the formula line is the messenger's speech, the sender's words one level deeper), and a speaker's command to say it (Isa 36:4: three levels); a psalm whose title ends with an inquit (Ps 18);
   • a message God tells someone to deliver ("דַּבֵּר אֶל־בְּנֵי יִשְׂרָאֵל", "וְאָמַרְתָּ אֲלֵיהֶם") is a quotation inside God's speech, spoken by the one told; in the law it runs to the end of God's speech, over chapter breaks (Lev 4:2–5:13, Lev 25:2–26:45, spoken by Moses);
   • the quotation marks of nine translations, as votes on the Hebrew words: Clear's per-word depth for eight translations and the BSB's own marks, compared within each quotation (whether a translation quotes a whole speech is its convention). Most of them nesting a stretch makes it a quotation inside the speech (a saying quoted in a speech, God's words in Moses' speeches, a letter, a vision's voice); a nested stretch of the layout that has neither a speech opener nor their support is dropped (a vocative, a "כִּי" clause of the same saying), and one that runs on past their support ends there.
4. The prophets: the frame is the book's prophet (prophets.tsv). Verse by verse, a stretch is God's words where "I" is God, and the prophet's own where "I" is the prophet (Jeremiah's laments) or God is named in the third person (Isa 11, 19); God's words that no quotation holds yet become one (basis oracle). A formula inside God's words ("נְאֻם־יְהוָה", "אָמַר יְהוָה", "לָכֵן כֹּה אָמַר יְהוָה") stays inside his quotation. Prose narrative outside any speech ends a stretch (Jer 36:4, Isa 37:36).
5. Speakers are checked against Glyssen's characters for each verse; nested speakers are read from the Hebrew inquit (God for a divine name, a person's name matched to Glyssen's characters, the outer speaker for "I"), else from Glyssen's other character for the verse, else from "I" through MACULA's participant references.
6. Hand corrections (corrections.tsv), each with its reason.
7. The review rows (review.tsv) last: changes from a review of doubtful passages, each with the reason the Hebrew gives (see below).
8. Names in Hebrew (shoresh/macula/hebrew_speakers.py): the steps above work with English labels (Clear's speakers, Glyssen's characters, the English names in prophets.tsv, corrections.tsv and review.tsv); none of them is in the data. Each quotation's speaker is read afresh from the Hebrew: the inquit is the verb of the nearest clause above the quotation's first word in MACULA's tree, else the nearest אָמַר or דִּבֶּר in the 20 words before it; its subject is MACULA's subject link (subjref, all of them: וַתַּעַן רָחֵל וְלֵאָה gives both), else the clause's subject. Where there is no inquit, a נְאֻם in the quotation's first words names the speaker. A subject that is a pronoun is not followed: MACULA's participant references agreed with the speaker data's speaker in about half of the cases, against three in four for a written subject (and most of the rest are the text being more exact: אַבְרָם before Gen 17, יִשְׂרָאֵל for Jacob, אָחִיהָ וְאִמָּהּ in Gen 24:55, הַמֶּלֶךְ), so an empty name is left rather than a doubtful one. A review row name gives the words that name the speaker where the subject does not (Ezra 1:2: כֹּרֶשׁ, not רוּחַ כֹּרֶשׁ). A quotation whose inquit names no one (God's words in the prophets, a psalm's voices) takes the Hebrew name of the same speaker from another quotation: the speaker data's label decides only who it is, the name and speaker_ref come from the text there. The name is the speaker's usual one in the book (most of that speaker's named quotations there, at least two), else in the whole Bible; it must be a proper name or God's name (not מֶלֶךְ, אִישׁ, עַם), God's name only for a divine speaker, and a name that is this speaker's rather than mostly another's ("false prophets" are not named יִרְמְיָהוּ). "Psalmist" (any untitled psalm) is never carried over. What is still unnamed keeps the label in alt_eng.
9. Characters (speaker_id, speaker_name): one id for each of the speaker data's characters (Clear's and Glyssen's labels, which already tell people apart: Pharaoh (1st) and (2nd), the two Lamechs), kept in a table in the repository so that ids do not change between releases (the labels in it are English, so the table is not part of this release). Two labels of one person (Abraham, Abraham (Abram); Samuel (boy), Samuel) share an id. A label without a proper name gets one id per book, so that unnamed people with the same label in different stories ("people", "someone") are not made one. The name is the proper name (or God's name) that more than half of the person's quotations named by a proper name, and at least a third of those named at all, are named by in their own inquit, and that no other person is named by more than three times as often (a messenger quoting כֹּה אָמַר שָׁאוּל, a prophet quoting יְהוָה); narrators get none. Where no inquit names a person the text names elsewhere, the table gives the word that does (רָחָב, Josh 2:1; חַוָּה, Gen 3:20; the prophets' headings): 50 people. Where the speaker data's label is the tradition's guess and not the text's (the brothers in Gen 37, the young Miriam of Exod 2), no name is given. The id and the quotation's own inquit agree: where the speaker data put a line in the wrong mouth (Rebekah's אֵלֵךְ under Laban, Gen 24:58; Saul's question to Abner under Abner, 1 Sam 17:55; God's words after כֹּה אָמַר יְהוָה under the prophet who delivers them, who is the quotation around them), a review row gives the right person; where the inquit names the wrong person (the quotation starts at its own verb, or "the days of David" for David), a review row gives the word that names the speaker. Where the inquit names several (וַתַּעַן רָחֵל וְלֵאָה) or a plural verb has several speak (the three visitors, Gen 18:9; the officials, Jer 36), the speaker is a group with its own id (one across the Bible where each member has a proper name, else one per book); speaker_members lists those of them who are characters of their own. The words a prophet is told to say (וְאָמַרְתָּ אֲלֵיהֶם) are his; the words after כֹּה אָמַר יְהוָה inside them are God's.

The code is shoresh/macula/build_speaker_layer.py in bcv-commons/bcv-query.

The review

48 doubtful passages (8 of each kind: the translations split, nesting from the layout alone, God's words the layer made a quotation, parallel texts that differ, unfoldingWord's notes where the layer doesn't nest, unnamed speakers) were reviewed by two AI agents independently (Claude Opus 5.5), each given the Hebrew of the passage with the evidence as numbers (the translations' votes, the grammar, Glyssen's characters, MACULA's participant references), no English text and no BHSA, and the conventions this layer follows. Where the two agreed, their rows were taken; where they differed, a third agent decided. Each row records its source (agent:claude-opus-5-5:review-v1, +adjudicated, +rekeyed where a row was moved to the build's current quotations) and a confidence. The layer is also built without them (--no-review). A review of all doubtful passages (about 700) is planned.

Since v0.11 the clause layout takes its speech levels from this layer, and mistakes found there are fixed here as more rows (source agent:claude-opus-5-5:layout-check-v1, each with its reason from the Hebrew): in v0.11, Lev 17:14-16 (the message of 17:8 runs to 17:16; "וָאֹמַר" is God's), Deuteronomy (Moses' first speech is one quotation, 1:6-4:40, and his second one, 5:1-26:19, with God's words inside them; 5:28-31, 9:13-14 and 18:17-20 are God's, not Moses'), Jer 1:7-8 and 2:2-3 ("נְאֻם־יְהוָה" closes God's words), Jer 24:5-10 (God's words end with the chapter) and Jer 25 (Jeremiah's speech of 25:3-38, with God's words inside it).

How it was checked

Against BHSA, privately. The rules were developed and tuned with the help of a comparison with the ETCBC's BHSA (CC BY-NC 4.0), made privately by bcv-commons/bsb-add-books as a reference for evaluation only: only figures and failure cases in words were exchanged, and no BHSA values, codes or files are in the code or in this release. BHSA marks each clause as narrative, quotation or discourse, with nesting, and leaves many clauses undetermined (most of Ps 19, 107, 125, 127 and 133, for example).

BHSA counts a text that is speech from its start (a psalm, the laws of Exodus and Deuteronomy, many oracles) as a quotation at its top level; this layer records that as the frame field. So the figures are given three ways: A with BHSA's top level counted, B without it, and A with the frame counted as a level. Measured on v0.9 (before the formula and review fixes below), per clause segment in or out of a quotation:

Genre	A	B	A, frame counted
All	82.1%	86.4%	91.8%
Narrative	97.9%	93.5%	98.4%
Law	87.2%	63.8%	87.2%
Prophecy	77.1%	74.2%	89.0%
Poetry	15.0%	96.9%	84.9%
Wisdom	66.6%	84.8%	66.6%

Quotations inside quotations (v0.9, by the layer's own levels): of BHSA's nested segments the layer has 54% (narrative 74%, law 88%, prophecy 40%, wisdom 43%); of the layer's, 88% are nested in BHSA too.

Quotation edges (v0.9): of the layer's starts, the share at a start in BHSA, and of BHSA's starts, the share the layer has; the same for ends:

Genre	Starts	Ends
Narrative	97% / 92%	96% / 93%
Law	89% / 86%	69% / 75%
Prophecy	78% / 75%	54% / 62%

The prophecy ends were low because an oracle formula inside God's words ended his quotation; this release keeps such formulas inside it, which bsb-add-books' check traced most of those ends to.

The prophets in both readings (v0.9, the share of segments at BHSA's exact level): levels alone 49%, the frame counted 46%, and the better of the two per segment 83% (v0.8: 45%). BHSA itself follows one convention in some books (levels alone in Ezekiel, Obadiah and Malachi; the frame counted in Micah, Nahum and Habakkuk) and both within others (Isaiah, Jeremiah, Zechariah). The levels are right in most of the prophets once the reading is chosen.

The review rows (v0.9 pilot): of the segments they changed, 163 moved closer to BHSA's level and 39 farther; 21 of the 39 were one row applied to the wrong quotation (2 Kgs 19:29–34), fixed in this release, and 12 a reading by convention (Amos 3:3–5, the prophet's own questions).

Our own measures for this release (bcv-query's checks):

• unfoldingWord's translationNotes mark 647 quotations inside quotations in the OT (figs-quotesinquotes, CC BY-SA 4.0; used to measure, not as input). With the frame counted as a level, the layer nests at 87.0% of them (narrative 87.5%, law 85.5%, prophecy 87.4%, wisdom 89.7%); without the review rows 84.9% (v0.7: 54%).
• Parallel texts, aligned word by word: 77.4% of matched words at the same level (the frame counted) and 93.4% with the same speaker (the English labels of v0.11); with the Hebrew names of v0.12 90.5%, as the two texts at times name one speaker differently (הַמֶּלֶךְ for דָּוִד). Some differences are by design (the Decalogue in Exod 20 and in Moses' speech, Deut 5).

Known gaps (open work):

• the Rabshakeh's words in Isa 36:7b are one level deeper than in 2 Kgs 18:22;
• a decree quoted in a letter (Ezra 6:3–12) and most levels of Nebuchadnezzar's letter (Dan 4);
• a nested speaker named through "I" can be the outer speaker where the reference chain does not reach the quoted person (Jer 22:18);
• the prophecy ends have not been measured against BHSA since the formula fix;
• the review covered 48 passages of about 700 doubtful ones;
• speaker_id comes from the speaker data's labels, checked against the inquit where it names one person by a proper name or several people; a wrong label on a quotation without such an inquit can still put a line in the wrong mouth, and one person under two labels the build does not know as one has two ids;
• 363 quotations have no Hebrew name: no inquit before them (most of them in poetry and the prophets), or a subject that is a pronoun or only in the verb; 251 of them keep the English label in alt_eng.

Sources and licence

Part	Source	Licence
Quotations (and their speakers, used in the build only)	Clear's "MACULA Quotation and Speaker Data", © 2023 Clear Bible, Inc.	CC BY 4.0
Character names per verse (used in the build only)	Glyssen CharacterVerse.txt, © 2014–2020 SIL LSDev and Faith Comes By Hearing	MIT
Who speaks (speaker, frame)	MACULA Hebrew's subject links and lemmas (lowfat, SHA-256 a8787f7edfc3…)	CC BY 4.0
Hebrew words, syntax	MACULA Hebrew (Clear Bible / Biblica)	CC BY 4.0
Clause segments	Hebrew Bible Clause Layout (bcv-commons)	CC BY-SA 4.0 (the segments are referenced by key)
Names of persons (English, used in the build only)	STEPBible TIPNR (Tyndale House)	CC BY 4.0
Quotation depth per word in eight translations (numbers only)	Clear's alignment-quotations, Clear-Bible/speaker-quotations @ b09e308a3a1a	CC BY 4.0
The BSB's quotation marks per Hebrew word	Berean Standard Bible Translation Tables	public domain
Participant references	MACULA Hebrew (Clear Bible / Biblica)	CC BY 4.0
Review rows	written by AI agents (Claude Opus 5.5) from the Hebrew and the evidence above, under bcv-commons' conventions	part of this layer
This layer	bcv-commons	CC BY 4.0

Attribution when you use this dataset:

Hebrew Bible Speaker Layer (v0.15), bcv-commons, CC BY 4.0. Quotations and speakers: MACULA Quotation and Speaker Data, Clear Bible, Inc. (CC BY 4.0). Character names: Glyssen, SIL / FCBH (MIT). Syntax: MACULA Hebrew, Clear Bible / Biblica (CC BY 4.0).

Corrections and feedback

Hebrew readers are invited to refine the layer: a correction is a row in corrections.tsv (the quotation's first word, the field, the new value, and the reason the Hebrew text gives). Corrections to the review rows are welcome the same way. Issues and requests: bcv-commons/bcv-query.