Ruach Studio 2.0.8

In honour of the whole Studio at once: 2.0.8 is the all-in-one release. The Writer is brought into order and opens two worlds, the Musician's and the Narrator's, each with its own tasks, templates, charter and preferences, and its Writing room keeps every answer, guards your lyrics and stops when you say so. The Narrator grows: Higgs TTS 3 from Boson AI reads beside Drama 3 as its second engine, figures are said in words in the case the sentence asks, the stresses of Russian and Belarusian come with it beside Ukrainian's, polytonic Greek is read as monotonic, a seam guard and a de-clicker take out the knock where pieces meet, a voice builder makes voices from a description, and twelve stock voices of Drama 3's own come with the Studio. The Videographer gets two viewers, as an editor has, and the edit's own tracks. Every model the Studio takes is kept at home by its content and asked at its pin before each cut; long songs are sung whole by default, the pieces an experimental choice; and the engine was measured on Vulkan beside SYCL on the Intel Arc, where SYCL stays the default. And every room is walked through under Viktor's law of compactness: its text buttons drawn as icons with their words in the tips, one height for a field and one for a bar, the Studio's own picker for long lists (HERESY 2392–2604).

New

• The Writer in two worlds, the Musician's and the Narrator's (HERESY 2486, 2487, 2497): with a reading open, the Writing room turns to the Narrator's world: its own tasks (Chapter, Scene, Dialogue by roles, Narration, Prepare the text for reading, Review & polish the reading), its chips (Stresses, Numbers in words, Pauses and breath, Roles and voices, Feelings), its templates, charter and draft, and «Your writing preferences» are that world's own, kept by the lab with their versions and carried by Export and Import; with anything else it is the Musician's world as before. A document moves between the worlds: the Narrator's «Edit in the Writer» brings a reading back as a version or a new document, a one-time sorting offers the old documents as readings (one with lyric section tags is never ticked for you), and a reading's takes link back to the document they were read from, by the lab itself.
• The Writer brought into order (HERESY 2490, 2494, 2495): the bar carries the Library's realm knob (Musician · both · Narrator) and the notebook's acts as icons; with no document open the right side is a canvas of action cards (New song, New reading, From the Musician, the Musician's versions, Trash, Import, Export) above a short guide; the document, the Writing room and the LLM assistant's output are three steps of one accordion, one open at a time, so nothing needs a scroll; the documents' column holds 360–420 px; the document's head keeps its way out in words (Load into the Musician for a song, Read in the Narrator for a reading) and the rest as icons; who writes, with which model, temperature and length, folds into one line that says them.
• The Writing room keeps your words and every answer (HERESY 2465, 2466, 2467, 2471, 2481): it writes into the document open in the notebook, the version before kept, or into the Musician's song when that is the one in hand, and a line above the button says which; a draft asked for the style only never brings lyrics in, and any other draft shows the lines of your lyrics it changed, old above new, and asks before Apply takes them; the structure template goes only with the tasks that write lyrics; an all-empty draft is said as nothing and changes nothing, and a field the model left alone is marked; every answer is kept with its document (its brief, task, model, time, tokens and raw answer), and any of them reopens as it was.
• Higgs TTS 3 from Boson AI, the Narrator's second engine (HERESY 2424): Engine → Models has a row «Narrator · Higgs TTS 3 (Boson AI)» that fetches the 4B model and its audio tokenizer from Boson's own repositories at their pins once you accept Boson's licences; Ruach Studio mirrors none of it. It reads in BF16 through transformers 5 in a Python of its own (tools/narrator-higgs-setup.sh makes it), on a CUDA card or the processor, one generation a piece, its codes kept; Boson's attribution stands in the Engine's About wherever the engine is here, and on every Higgs take (docs/NARRATOR-HIGGS.md says what the licences ask of you). A model behind a licence or a gate asks once, in one small dialog: its licences with their links, or the gate's link and a token field used for that one download and kept nowhere.
• The UIX pass: maximum compactness, minimum clutter of text buttons (HERESY 2526, 2527, 2530, 2541, 2567, 2580): Viktor's law of 12.11.2026, «Максимум компактности, минимум текстового кнопочного клаттера. Это закон UIX нашей студии», walked through every room by an audit of about a hundred findings: one set of icons for one act everywhere (Lucide's shapes, ISC, under Lucide's own names, about a hundred of them added), one height for a field (36 px) and one for a bar's control (30 px), one way into the Studio's own picker for every room's long lists, and a button's words kept in its label and its tip, so every language still reads them.
• The Narrator, the Musician's mirror (HERESY 2401, 2510, 2511, 2512, 2513, 2514, 2515): the room stands in the Musician's cards (the lectern, the voice, the engine) with Read in a bar at the foot, everything in sight at 1920 px, a speaking voice's envelope breathing under its name while a run reads, and its buttons as icons with their words in the tips. A reading's take is read again in the Narrator with every setting and a new seed, never sung by YuE2; a run card plays, regenerates, fixes and checks with Whisper, and pours its reading into the form four ways (everything but the seed, the voice and settings, exactly with its seed, only its text) with an undo; a reading's passport and its menu speak of a reading, not of a song.
• The Narrator says numbers in words, in their case (HERESY 2397): «на 6650 часах» goes to the engine as «на шести́ ты́сячах шестиста́х пяти́десяти часа́х», the case from the noun's ending and the preposition, the gender of 1 and 2 from the noun, years and «5-й» as ordinals, «%» as its noun; Ukrainian the same («на шести́ ти́сячах шестиста́х п’ятдесяти́ годи́нах»), Greek by the noun's gender («σε 6.650 ώρες»: «έξι χιλιάδες εξακόσιες πενήντα ώρες»). Times, verses, dates, versions and numbers glued to letters stay as written; Check shows what it said; Whisper's figures are read the same way, so a number in words is no error of the reading.
• Russian stresses come with the Narrator (HERESY 2399, 2458): every form of A. A. Zaliznyak's Grammatical Dictionary, unfolded by the Studio from gramdict's data (CC BY-NC 4.0, non-commercial use), 2.2 million forms with their stresses, the verbs whose persons the dictionary writes out (дать: дам, дашь, даст…) or names one by one (хо́чешь, хо́чет) taken as written; a proper name is looked up only for a capitalized word, so a country never hides a pronoun (То́го, того́).
• Stresses as the language says them (HERESY 2440, 2447, 2451, 2452, 2454, 2475): Russian reads gramdict's index first; a clitic group takes one stress as Russian says it (не́ было, не́ был, на́ пол, по́ полю, бе́з вести, не́ за что), за never, and your own mark on either word wins; a form the dictionaries stress two ways (за́мок and замо́к, му́ка and мука́) or twice within one word stays bare unless you mark it or your dictionary has it, so a dictionary never forces a stress it is not sure of; Drama 3 is handed Russian stresses as capital vowels and Ukrainian ones as a capital with the acute, the forms Viktor's ear chose.
• The stressed vowel as a capital (HERESY 2398): an engine may take the stress as the acute (as before), as a capital vowel, or both, per language (engine.json stress_lang); the Greek question mark ends a sentence and the ano teleia is a place to cut a long one.
• Belarusian stresses come with the Narrator (HERESY 2459): every word form of GrammarDB, the Belarusian NLP community's grammar database (CC BY-SA 4.0), 1.26 million forms with their stresses; a word with ў after a vowel is found by its у. On a Belarusian text stressed by hand it knows 286 of 303 words and agrees with 236 of the 244 it stresses one way.
• Polytonic Greek read as monotonic (HERESY 2460): Ἐν ἀρχῇ ἦν ὁ λόγος goes to the engine as Εν αρχή ην ο λόγος (the same sound today: the breathings and the iota subscript silent, every accent the stress, the lunate sigma σ or ς); the text in the editor stays as written.
• No knock where a reading's pieces meet (HERESY 2419, 2420, 2421): the short low knock between sentences was audio.cpp cutting a piece over 128 characters into generations joined with no fade; Drama 3 now reads each piece in one generation and keeps its codes beside it for a decoder of tomorrow; a seam guard finds the codec's drop where a reading is joined and mends only those milliseconds, the take's note saying how many; and the Speech de-clicker, the first node of the Narrator's chain in the Refiner, mends the takes already made as a branch, with a reel of each spot before and after.
• A voice builder and twelve stock voices (HERESY 2403, 2449, 2455, 2456): «Voice builder…» makes two to four candidate voices from a description (a gender, seeds and the words to read if you like) in one run of Drama 3, measures each (pitch, brightness, background, creak) and keeps the one you choose as a made-up voice of your own; twelve made-up voices come with the Studio, one for each of the Narrator's languages, six men and six women, made by Drama 3 itself from words alone (the Japanese one made again, calmer and lower), marked never for training, and they are Drama 3's own set: its timbre groups everywhere. Built with Fish Audio.
• The Narrator's ear: on, off, or afterwards (HERESY 2434): Hear each piece (Whisper) stays on by default; off, the take says it was not heard, never passing for a clean one; Check with Whisper hears a finished take afterwards, from the Library or its run card, piece by piece against the words it was asked to read, the words lost, added or swapped with their times beside the take and in its note, the audio untouched.
• Drama 3 on an Intel Arc, and Q8_0 by default (HERESY 2426, 2435, 2436, 2441): Drama 3 reads on an Intel Arc through audio.cpp's Vulkan build, which tools/narrator-setup.sh builds on a machine with an Intel card (the guide names the packages for Debian, Ubuntu, Fedora and Arch), with a tier of its own (Q4_K, 28 of 40 expert banks in RAM, RTF 3.3 on an Arc Pro B50), the Intel driver alone shown to the loader and the device read back by name, never the processor in the card's name; on NVIDIA cards Q8_0 is the default wherever it fits (Viktor's ear: 20 of 20), then Q6_K, Q4_K last, and a card of 24 GB fetches Q8_0 at setup.
• A book is one card in the Library (HERESY 2402): a reading in parts shows as one card (its name, its length, its parts and their likes summed), its parts off the grid; a click opens it over the page: the book played through or from part K, a strip of the parts by their length, and the parts as the Library's own cards in two columns.
• The Videographer's two viewers and the edit's own tracks (HERESY 2392, 2393, 2394): the animatic or the clip on the left, the edit's program always on the right, each as large as the room leaves (one viewer with a switch under 1600 px), and one sound in the room: whatever starts playing pauses every other player; the edit's tracks each its own (the shots, the titles in their template, the karaoke from the take's Whisper timing, the look, the song), an item chosen opening in the inspector; the titles and the karaoke in a font of your own, installed or a .ttf or .otf file, which the render really uses, never a silent fallback.
• Every model kept at home, every pin asked before each cut (HERESY 2423): every file of every pinned model the Studio takes (44 pins, ours and the others') is found by its content, never by its name, and kept in a cold copy proven by reading it back; before each release a guard asks Hugging Face live whether each pinned repository still answers, its commit still exists and each file the Studio takes is still there with its size and its sum, and a loss stops the cut, as does a Hugging Face that does not answer.
• Long songs whole by default (HERESY 2509): the sound half sings a song through one window of the model's context (about 6:16 to 7:52 on the whole context, by the length of the style, lyrics and score; less on a card with less memory); a song the model would sing past it is held whole at the window and the take says so; the pieces of 2.0.7, joined in a crossfade, are an experimental choice now, Sampling › Sound and output › Long songs, and a kept part or a source already past one window goes on in pieces as it was made.
• A score without its key line is sung, and its key offered (HERESY 2433, 2444): the engine sings such a score as written, with no key signature, and says so in its log, the take and the page, where it used to stop the run; the page reads it too and offers the key, or two, its notes and chords read as, with how sure, and puts the line into the score's header at a click.
• One rule for every name (HERESY 2462): wherever the Studio names or renames something (a workspace, a section, an album, a take's title, a preset, a profile, a mix, a bank, a voice…), a name never holds / \ * ~ ? : " < > | or an invisible control character and never begins with a dot; the dialog says what it refuses and offers a cleaned name, never changing one quietly, and the lab holds the same rule, so a script cannot go round the page. A name already there is not judged.
• The Trainer's eight charts, for every run (HERESY 2442): the loss, the sound half's flow loss, the music half's cross-entropy and its KL to the base, the learning rate, the gradient norm, seconds a step and each card's peak memory, in one order for every run, read from whatever the run logged; a metric a run never logged is an empty chart that says so, never a line at zero.
• The engine measured on Vulkan beside SYCL (HERESY 2450): built with ggml's Vulkan for the Intel Arc, every op of our graphs runs on the card and the sound matches SYCL's as closely as SYCL matches itself, but it sings 1.6 to 3.4 times slower, so the Arc's engine stays SYCL (./build.sh --sycl); the engine gains --no-fa-nar (no flash attention in the sound half) and YUE2_GGML_DEBUG=1 (ggml's own debug log let through).

The Narrator

• Drama 3 hears a paragraph's mood through all its pieces: a piece that does not open with a mood of its own reads in the one said last before it in its paragraph; a sound in its place ([sighs]) or a pause is not carried, and a new paragraph starts clean (HERESY 2439).
• Two trials for the emotional arc of a long reading, off unless a run asks: each piece read with the one before as its history, and longer pieces (HERESY 2474).
• The Narrator reads Japanese and Chinese in whole sentences, cut at 。！？ (HERESY 2427).
• «!».» and «?».» no longer leave Drama 3 silent after them, and a piece the engine answers with an error is asked again with another seed, three times in all, before the run fails with the engine's own word (HERESY 2422).
• The speech chain's denoiser no longer leaves a take 25 ms late: the result lies sample for sample where the take was (HERESY 2418).
• The Narrator's takes land in the Narrator's realm, a workspace named for a reading too, and a speech file imported with realm=reader is the Narrator's (HERESY 2437).
• A reading just done is never written over as «the lab restarted» (HERESY 2446).
• Drama 3 on the Intel Arc reads in audio.cpp's streaming session, its codec decoded four frames at a time as they come: the same sound (the two decodes 51 dB apart), 9.1 GB at the card's peak instead of 15.1 on a long piece; NVIDIA cards and the processor read as before until they are measured (HERESY 2516).
• The Narrator's guide says the Ukrainian and Russian dictionaries come with the Studio, and that any other language plugs in its own (HERESY 2429).
• Notes take 6000 characters, every one alike, a take's own too: a longer one is cut at 6000 and ends in «…», never silently; a tooltip shows a note's first 1200 (HERESY 2404, 2438).
• The Narrator's room in the window: the go bar under the form, never over it, the voice and the engine in a column of their own, the room's words in an (i); a shut run card shows its head and two whole lines; its fields on the Studio's height (HERESY 2560, 2561, 2575).
• The Narrator's quant is the Studio's picker, its rows in the page's language; the tier line said whole; the engine's first-run words in the (i) of its Quant (HERESY 2562, 2563, 2564).
• Where a run read and its seed in the page's language, the passports' heads in icons; what the Narrator writes into the page follows a change of language, and the palette's 27 tags have their tips (HERESY 2565, 2566).

The Writer

• The voice builder's language, gender and number of candidates, and the Writer's «Compare with», say what they do in their tips (HERESY 2600).
• The Studio's own writer on the Intel Arc: Engine → Writer and Artwork asks where it runs (Auto, a CUDA card, the Arc or the processor) and, on the Arc, SYCL (the default, quicker to answer) or Vulkan where its build is there; it waits for its turn there, never falls to the processor behind your back, and leaves the card the moment a song, the Narrator or the watchdog needs it (HERESY 2463).
• The Studio's own writer reads 32,768 tokens; each question is counted first with the writer's own template and tokenizer, the room says the input and what is left for the answer before it comes, and a question that leaves less than 512 is refused with its numbers (HERESY 2464).
• The Studio's own writer answers the Writing room by a JSON schema, so a small model in a narrow window no longer answers half an object (HERESY 2476).
• Stop beside Send (or Esc) aborts a request and tells the lab, which stops the writer; a wait that runs out says «No answer from … in 15 minutes», never a word about CORS (HERESY 2469).
• The Narrator's transcript cleanup reads the Writer's answer again; it had ended «the model gave nothing» every time (HERESY 2468).
• The Writing room's words are all in the page's language, its line names the writer by its model and where it runs, and its pickers are as wide as their longest choice (HERESY 2470).
• The OpenRouter key may stay in this browser tab's session only, never written to a file; saved, it lives in the Studio's user folder as before (HERESY 2473).
• A document's version against now, and a document against any of its versions or another version of the same song, line by line (HERESY 2477, 2482).
• The lyrics meter on the notebook's LYRICS and the Writing room's draft, as the Musician has it (HERESY 2478).
• The notebook's list puts a song's versions under one name (by the version in its title), sorts and filters (songs, readings, with takes), brings the trash back with Restore, and shows a short guide where the cards were (HERESY 2479).
• A document is a song (10,000 characters, said past 9,000) or a reading (no limit); a paste past the limit asks: keep it whole as a reading, cut it, or cancel; the Musician's lyrics say how much a long paste lost (HERESY 2480).
• A big cut in the notebook is never saved silently: it asks to keep it as a new version, undo it or keep editing (HERESY 2492).
• A take links to the document in hand only when it is the same song, and one linked by mistake comes off by an icon (HERESY 2493).
• The Writer's STYLE gets the Musician's instrument chips and cheat-sheet, and the saved styles, from the same modules (HERESY 2491, 2498).
• The LLM assistant's step says how many answers the song has kept (HERESY 2496).
• «aa» and «AA» beside the Musician's lyrics and the Writer's and the Narrator's text: lower the selection or the whole text, raise the selection, the tags in brackets never touched, one Ctrl+Z to give it back; Shift+click spares a capital that carries a stress mark (HERESY 2453, 2472).
• The Writer's order, filter and search in its bar as icons; its document's acts as icons, its profile in the picker, its fields one height (HERESY 2568, 2569).
• The page's words in the page's language: Models and LoRAs, the default VAE's tip, the Writer's dates; «the Studio» with its capital in the slow line's tip (HERESY 2532, 2534).
• The take's menu drawn in the Studio's icons (HERESY 2540); the Writer's card and the pick buttons on the Studio's scale (HERESY 2542); the Musician's text and glyph buttons, the case buttons and a LoRA of another model drawn (HERESY 2543, 2547).

The Videographer

• A finished clip heavier than 6 Mbit/s gets a light copy (720p, under 3 Mbit/s) for a slow link: «Light copy» plays it, and two stalls within 20 s switch to it by themselves; the download stays the master (HERESY 2395).
• The Videographer's inspector scrolls beside the timeline, so the program monitor keeps its size when a title is chosen (HERESY 2431); a take chosen again reads its lyrics' timing anew, so the karaoke fills once Whisper has timed it (HERESY 2432).
• The Videographer's titles track is called ST1, the text-track name an editor knows from Resolve (HERESY 2517).
• The Videographer: the viewers' player bar a row of its own under the picture; the shots', the bar's and the edit's buttons as icons with tips; its lists as the Studio's picker, «Animatic and clip» in one row (HERESY 2587, 2588, 2589, 2592, 2595).
• The Videographer with no take or no plan says so in its middle with the next step; a failure's chip gives the reason through the catalog, the lab's words in its tip; a shot's fields as tall as their words; the find field's icon and its three tips in the page's language (HERESY 2590, 2591, 2593, 2594).

The Artist

• The Artist: a picture's download and «Set as cover» two icons over its corner, the library's column never wider than itself; a run's three acts as icons; the board's schedules and a failure's raw words on one line; its head in one line, its words in an (i) (HERESY 2581, 2582, 2583, 2584).
• The Artist's lists in the Studio's picker, their tips in the page's language; the form's words and glyphs on buttons as icons (HERESY 2585, 2586).

The rooms and the house

• The Engine opens on Ctrl+Alt+9 (Ctrl+9 in the installed app); Ctrl+0 is never taken, so the window's zoom back to 100 % always works; the rooms' tips say their keys as the keys open them (HERESY 2405).
• In the playlist every take starts from its beginning; outside it the player still comes back to where it left a take (HERESY 2406).
• A dry bounce carries its lanes' room and delay onto the new lane's sends, so the mix keeps them and they stay live (HERESY 2407).
• In Hebrew the compare window's transport runs left to right, as the player's (HERESY 2408); a tip says why a section's word is translated inside a sentence while the tags stay English (HERESY 2409); the Videographer's field reads «Bible verse» (HERESY 2410).
• Eight new demo songs, one each in Ukrainian, Belarusian, Greek, Spanish, Italian, French, Portuguese and German (HERESY 2412).
• A Hebrew page reads the English guide left to right until the Hebrew one comes (HERESY 2413).
• The Engine's About credits Resemble Enhance only where it is installed (HERESY 2415).
• The bar names a training run as it is: «Training» for a run that teaches no adapter, «Run» while it runs, and each card of a run held on several, never «GPUNaN» (HERESY 2425).
• The Library's Archived row folds as a workspace's sections do (HERESY 2445).
• Abbreviations and names stay themselves in every language: LUT, EQ, GPU, VRAM, RAM, CPU, LoRA and the others, and «Whisper», the model's name (its voice chip is «Whispered voice»); ordinary words stay translated (HERESY 2430, 2443).
• Russian keeps the root «раб» out of what the Studio says, as Ukrainian and Belarusian do: a machine that works «работает», while the language model and the rooms create (HERESY 2488).
• check-i18n's Hebrew rule counts the vowel and cantillation marks, not the maqaf (HERESY 2428).
• Every room's tab has its tip, the Artist's in the page's language (HERESY 2521).
• The folded log is a button in the player's row and a training lives in the system bar: neither lies over a room's fields any more (HERESY 2522); the Refiner's room is no longer 43 px taller than the window (HERESY 2533).
• The system bar fits at 1920 px, a job its kind and its percent, every disk in sight (HERESY 2523); the empty player is one row, and the one-row player gives its title the width before its small wave (HERESY 2524).
• A toast steps under an open menu; the bar's right group stands at one height; the Engine's access line and About in the page's language (HERESY 2525).
• The Engine: Models and LoRAs' «have it» and «fetch» as icons, its cards' one-act buttons as icons, every tab's cards from the top and as wide as the grid (HERESY 2528, 2529, 2531).
• The Musician's long lists (the melody's take, the song's key, the imported MIDI's key) in the Studio's picker with a Find; its title, version and seeds at the Studio's height (HERESY 2544, 2549).
• A run's card keeps every value on one line, cut with an ellipsis and whole in its tip; the shape's ranks and the note fit their card (HERESY 2545, 2546).
• The Librarian's glyphs drawn and its words in Russian; a song's passport says its mode and its decoder in the page's language; no English left in the Musician's tips (HERESY 2550, 2551, 2552).
• The Refiner's bar one height, the bucket an icon with its count, a pipeline's plugin named once (HERESY 2570).
• The Trainer's head: its words in an (i), the guide and Reload as icons, the steps with tips; the charts' and the runs' glyphs drawn, «Compare with…» over all the runs in the picker with a Find; «Reading the runs…» while the list comes (HERESY 2571, 2572, 2573).
• Thirty new words in the twelve catalogs for the pass (HERESY 2548, 2574).

The engine, the install and the AppImage

• An engine started with our decoders alone refuses to encode, with the reason, instead of crashing on a recording; /props says whether it has an encoder (HERESY 2484).
• One set of install sizes on the visiting card, in README, INSTALL, the launcher and the Narrator's guide: the install 22.5 to 25.9 GB, every room 122 to 240 GB (HERESY 2485); the card's «What comes next» says what 2.0.8 brings (HERESY 2414).
• The AppImage says how many GB its first install downloads before anything comes, stops honestly when the network goes and goes on by itself when it is back (HERESY 2411).
• The AppImage carries no patch leftovers (.orig, .rej): a packaging patch is applied only whole (HERESY 2416); nor the bytecode of the machine that packed it (HERESY 2601).

Downloads

• The source tarball and the tag, as every release: git pull (or Pinokio's Update) brings an installed Studio to 2.0.8, and Engine → Updates offers it.
• The Linux AppImage, RuachStudio-2.0.8-core-x86_64.AppImage with its .sha256 and .zsync: a window of its own first, offline too, its own C library, and the size of its first install said before anything downloads (README, «Linux: AppImage»).
• Higgs TTS 3 is not in either: Engine → Models fetches it from Boson AI's own repositories once you accept Boson's licences (docs/NARRATOR-HIGGS.md); Ruach Studio mirrors none of it.

Thanks

To m-a-p for YuE2; to ServeurpersoCom for yue2.cpp and ggml-org for ggml; to Fish Audio for Drama 3 (Built with Fish Audio) and to the authors of audio.cpp; to Boson AI for Higgs TTS 3 (Built with Higgs TTS 3 licensed from Boson AI USA, Inc.; Built with Higgs Materials licensed from Boson AI USA, Inc., Copyright Boson AI USA, Inc., All Rights Reserved and Meta Llama 3 licensed under the Meta Llama 3 Community License, Copyright Meta Platforms, Inc., All Rights Reserved) and to multimodalart for its transformers port; to gramdict for A. A. Zaliznyak's Grammatical Dictionary as data, to the Belarusian NLP community for GrammarDB, to lang-uk for ukrainian-word-stress and to ULIF for «Словники України», the Narrator's stresses; to everyone who opens an issue or a pull request.

Everything, entry by entry: CHANGELOG.md and heresy/docs/HERESY.md.