Parseh a0.4.4

Added

• Parseh lives in the parseh-io organisation on GitHub; Settings → Updating Parseh looks in the new repository
• The guide is at parseh.io/guide, under a bar that leads to Parseh's site; pages exported from Parseh link to it for good
• The phone app's icons are published with the guide again
• Every address Parseh writes for itself is kept in one place, lib/project.py
• release.py links asks every address Parseh names and says what each answers
• Settings → About shows a few friendly numbers: cards answered, videos and books on the shelves, since when
• From a0.4.2: prompts of your own, the studio's prompt in parts, IPA and short-vowel options, skills for your chatbot, a book made by an agent from any device, pictograms (ARASAAC)
• From a0.4.2: a video's vocabulary in the books' entries, notes on a video's phrases, a sound in place of a video, flashcards that turn
• From a0.4.3: Whisper models for each language, optional exact word times kept through every edit of a transcript
• From a0.4.3: a transcript workspace with LM likelihood, an LLM connection, a reasoning workspace and chatbot copy-and-paste review
• From a0.4.3: partial-word colours in the studio, mobile arrow keys that skip, a card maker for books and videos

Changed

• Speech to text lists the Whisper program first, above the second pass and the models
• a0.4.2 and a0.4.3 were finished but not released on their own: their full entries follow

Fixed

• LLM Integration, LM likelihood and About are drawn with their own stylesheet and end in a foot like the other doors
• The README and the settings pages link to guide pages that exist

[a0.4.3] - 2026-10-05

Added

• A standalone reasoning-workspace CSV checker, shipped with the helper and skill, for draft checks, exact source spans and required-entry coverage
• Standard and language-specific Whisper models in one pinned catalogue, with independent resumable installs, per-language model choices and model provenance in transcription jobs
• Optional Hugging Face downloads for Persian fast/accuracy, Arabic, Italian, Hindi fast and Spanish, including merged LoRA models, matching offline assets and original licence notices
• Settings → About with version, author, launch-script locations and installation information using the host's native paths
• Saved pending transcription reviews, with pause/resume from Videos across server and browser restarts
• Checked-word locks, selected-section review, single-word LM likelihood and bulk selection of the latest method's best choices
• Optional similar-sound and candidate search-probability filters for additional model-derived words; every genuine Whisper alternative remains eligible
• LM likelihood review with installed-model selection, raw-token candidate search, numerical rankings and an isolated CPU/GPU scoring worker
• External chatbot copy/paste review for suspect words, whole text and reasoning workspaces, with bounded prompts, validated answer import and downloadable text/CSV workspace files
• Additional reasoning workspace review with numbered transcript blanks, CSV evidence, isolated Python tools, separate model selection and two-pass word progress; existing review modes and thresholds stay unchanged
• Browser transcription workspace with video beside the transcript, word/caption replay, ±1/2/5-second playback controls, caption filters and a persistent word editor
• Brief word playback with automatic pause, installed-dictionary suspect flags and original/replacement meanings in browser transcript review
• Dictionary meanings for native Whisper alternatives alongside their recognition scores in the word inspector
• Native faster-whisper beam alternatives mapped to source words, with sequence log scores kept separate from word probabilities
• Browser Settings → LLM Integration: a reusable OpenAI-compatible endpoint, host-local credentials, model discovery and Unsloth run-settings link import
• Review-first Whisper results with short sentence LLM replies, word-count progress, local failure recovery, retry of unresolved words, keyboard edits and inspectable responses
• Remote installed-model profiles from Studio links, with explicit API loading of quantization, KV cache, context and vision options
• Separate models, prompts and downloadable skills for suspect-word and whole-text review, including confident words and contiguous ASR spans
• Individual pending draft decisions and explicit Use this transcript preserve the existing box and timing guards
• Optional, local exact word times for every speech-to-text language, with hash-pinned int8 CTC networks and a remembered add-page switch
• A Parseh-made transcript carries its word timing through the add-page editor and stores it with the matching video
• Two transcript tidy cues: the existing text cues and recorded pauses
• Mobile-mode keyboard arrows skip narration or video by the selected amount
• Partial-word foreground colours in Studio, with selection-first editing, Unicode-safe boundaries, one-word linguistic clouds, PDF and standalone-HTML output

Changed

• Transcript word editors can include adjacent words for a short-span correction; reasoning workspaces can join split words across suspect/confident neighbors
• Transcription opens directly into a shared editing workspace; correction methods are reusable tools, with optional automatic and later per-word/section Whisper rechecks
• All four transcript-correction methods are labeled experimental; LM likelihood offers no-cutoff, suggested 0.1 and custom cutoff choices
• LLM settings and transcript review show simpler controls, with technical setup and numerical diagnostics folded away
• Markdown and exercise-deck cards made from books or videos mark target and context fields as the target language automatically
• Studio and exported word clouds are linguistic only; foreground colour is applied from the source editor

Fixed

• Speech program and model downloads no longer fail before their progress starts
• Exported guides include the segmented-colour parser needed for independent builds
• Speech model notices list every catalogue source and converted package with its own licence
• Persian dictionary lookup recognizes colloquial copulas and combined nominal endings, with direct entries preferred and dictionary-backed stem guards
• LM likelihood candidate search preserves source whitespace with SentencePiece tokenizers, including PersianMind
• Partial-word colours compile and remain visible in PDFs, including headings, target-language blocks and exercise prompts

[a0.4.2] - 2026-09-30

Added

• Glosses an LLM writes are aligned to their chunks: a meaning says what its own words say, in the text's order
• Every prompt is assembled by one kit from its instructions, its answer contract and its data, and opens with a version line
• The LLM row: one control for every prompt, saying its size before the copy, on the studio, the exercise dialog, the add page, the tidy, the player, the reader and Ask LLM
• A video's vocabulary in the books' entries, asked of the chatbots and drawn as in the reader, with macro buttons that say what they write
• A book made by an agent in place, from any device, its text given in parts, watched from the library, steered through asks, finished with one button
• Prompts of your own, from a menu at every copy button, exported and imported under Settings → Your prompts
• The studio's prompt in parts you tick, with presets, a level and a length, and the exercise dialog pre-ticking what the page uses
• Options of a prompt: the transliteration in IPA, and the short vowels of Persian and Arabic
• Every language's conventions carry a worked example and one vocabulary convention
• Skills for your chatbot: gloss, markdown and book, built on request and downloaded under Settings → Skills
• A sound in place of a video, with its waveform
• A video phrase's note can be edited, and a ✱ notes button lights every phrase that has one
• A vocabulary or opposites flashcard's example, notes and source go on the side the card turns to
• Settings → Pictograms (ARASAAC): the symbols and their words fetched on request, with their licence

Changed

• The LLM prompt page's Edit prompt, Save custom prompt and Reset to default give way to the prompt menu