Repository navigation
BigLinux TTS 4.1.0: reliable History, Kokoro sound fix, release cleanup - #61
Merged
Merged
Conversation
- clippy -D warnings passes: the FFI module carries one documented allow(unsafe_code); the 100-line NFD match became a lookup table (with a unit test); the unused sample_rate() is gone. - espeak synthesis takes the sample buffer instead of cloning it, so the largest synthesis no longer stays allocated: 24 MB resident after a 41k-character text (was 133 MB). - Version 4.1.0 and the authors in Cargo.toml.
…, text rules per voice Reading - History records every reading that was heard, for every engine (RHVoice and Kokoro were never recorded), off the GTK thread; nothing when off or stopped. Entries are newest first, save_text=False keeps no text, the folder is private (0700), SQLite connections are closed (one fd leak per saved reading). - espeak-ng streams long texts in chunks like Piper (20k chars: first audio 0.04 s, was 1.37 s); its fallback reads the text from stdin (no option injection, no 128 KiB limit). One streaming routine for both engines; a chunk file lives until its player ends, at most two at a time. - A player started after Stop is terminated; simultaneous readings are tracked so Stop and Pause reach all of them. - MPRIS replay, the shortcut, the tray and "Test voice" share one set of options: the character limit and text settings apply everywhere. - SIGTERM/SIGINT/SIGHUP quit cleanly; koko's text never reaches the logs. Text - Rules follow the voice's language, not the system's. - Abbreviations: no single letters or real words, not inside e-mails/URLs, titles drop their period; one precompiled pattern per language. - Feminine hours, seconds; HTML only removes real tags; numbered lists and paragraph breaks kept; "www." URLs removed, or read as their domain. - The character limit cuts at a sentence or word boundary. Cleanup - Removed the unreachable speech-dispatcher backend (settings move to RHVoice), the PyTorch Kokoro path (lang_code bug), the khotkeys file and chain, launcher pinning and other dead code (~1300 lines). Interface and translations - Sidebar selectors show the value as subtitle (no more "P…"); icons that exist in Adwaita, Breeze and BigLinux themes; readable history badges and voice names; the tray holds a steady icon under reduced motion. - i18n follows GNU gettext precedence with per-message fallback; language and country names come from iso-codes; every string is a literal that the pipeline's xgettext extracts. Version 4.1.0; authors Rafael Ruscher, Bruno Gonçalves, Tales A. Mendonça.
Every catalog has all 322 strings. pt corrected to European Portuguese. scripts/sync_translations.py extracts like the CI (plain xgettext), keeps existing translations, and --check validates placeholders.
- epoch=1 pkgver=4.1.0 pkgrel=1 (26.10.08 > 4.1.0 for vercmp without the epoch); prepare() refuses mismatched versions. - Dependencies only from core/extra: wl-clipboard (wl-clipboard-rs provides it), alsa-lib added; BigLinux-only packages are optdepends. Built, installed, run and removed in a clean Arch container; upgrade from 26.10.08 keeps user settings. - tts_engine.so in /usr/lib/tts-biglinux (namcap); bytecode compiled in package() with installed paths. - ci.yml: Python, Rust and PKGBUILD checks in an Arch container. The translation workflow pins actions, limits permissions and tmate, and discards catalogs that fail the check. - LICENSE (GPL-3.0-or-later); ruff rules pinned.
Installation without the BigLinux repositories, dependencies, i18n, tests, CI, release process, third-party licenses, authors; new screenshots of the real application. The benchmark script finds the engine again.
Why History looked empty or lost entries:
- RHVoice and Kokoro readings were never recorded, and Piper/espeak-ng long
readings had no audio: only one code path saved history.
- A database migration and WAL switch raced between the save worker and the
History page ("duplicate column", "database is locked"): saves were lost.
- The Kokoro engine (`koko pipe`) played by itself, so its audio could not
be kept.
Now each speak() is a reading (_Reading) that owns its processes, files,
threads and recording, so simultaneous and queued readings never mix, and
_finish_reading is the one place that records it (completed / stopped /
error), only if it was heard, in a worker (first audio measured equal with
history on and off).
Audio kept is the audio played: RHVoice and `koko stream` write WAV to
stdout and a relay passes the same bytes to aplay and to the recording;
espeak-ng/Piper chunk files are appended as they start playing. Recordings
go straight into the history folder (.rec-…part) and are finalized there.
history_db: schema 2 (processed text, status, duration, audio file) added
in place; filters, full-text search, paging; serialized migration; JSON
migration without duplicates. history_service: atomic saves that never
overwrite; delete to the Trash; repair_history rebuilds a damaged index
from the files (kept aside), recovers orphan files and crash recordings,
fills durations, marks missing audio; never deletes a recording.
Tests use real processes (fake koko/RHVoice/aplay) and can no longer touch
the user's history, Trash or settings (autouse + session isolation).
… at start With the tray icon off, the state handler returned early: queued readings never played, notifications and the cold-start hold were skipped (since 4.0.0). History is repaired in a worker two seconds after start. Tests: shortcut/tray, media Play key and queue mode all reach the History.
…ettings - The main menu showed History only while "Save history" was on, and saving was off by default: for most people the History never appeared. It is now always available; turning saving off keeps every saved reading and a banner says new ones are not kept. - Saving is on for new installations; settings from earlier versions keep their value (a file without the key stays off, never turned on silently). - Advanced options: Keep the audio, Keep the text, Keep at most (retention), and where it is stored (only on this computer). - The page: cards with expandable text, engine, voice, date, duration, outcome (stopped / error), player with the known duration, copy, read again, save audio as, open the folder; search by text, voice, engine or date; date and engine filters; newest/oldest; list or grid; selection to save to a folder or delete; pages of 50 loaded on scroll, in a worker; refreshes itself when a reading is saved; clear empty states. - Deleting: one entry with Undo, a selection after confirmation (it used to delete everything at once, without asking, and permanently); files go to the Trash.
64 new strings (History page, history options, outcomes). pt-BR and pt by hand; the others validated for placeholders and line breaks. "History" corrected to the desktop term in hu, es, it, el and is (it was translated as history in the sense of past events).
… deleting New History and Advanced options screenshots; test count.
The package still builds biglinux/tts-biglinux by default (BigLinux build service); testers can install a version that is not published yet.
History: every reading recorded with its audio, always in the menu, recovery (includes 4.1.0)
koko stream writes its log lines ("Detected language…", "CALLING
PHONEMIZE ON…") into stdout among the WAV bytes: the audio relayed to
aplay was garbage and nothing was heard. koko pipe plays by itself and
writes everything it plays to -o as it goes, so that file is the history
recording itself (or a per-reading scratch file, deleted afterwards).
Sound start comes from its "Streaming audio" stderr line.
koko synthesizes ahead of playback: a stopped reading keeps only the time
actually heard (pauses not counted).
fix(kokoro): sound again — koko pipe -o instead of koko stream
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Brings the fork's main to upstream: three merged fork PRs (ruscher/tts-biglinux #9, #10, #11).
4.1.0 release (#9)
epoch=1,pkgver=4.1.0) and README. A test keeps them in sync.TTS_BIGLINUX_SOURCEcan build another repository or branch.Reliable History (#10)
Kokoro sound (#11)
koko streammixes its log lines into the audio on stdout, so Kokoro was silent.koko pipe -o FILE: koko plays the audio itself and writes it to FILE, which is the History recording. Sound start is detected from itsStreaming audiostderr line.pkgrel=2.Verification
1:4.1.0-2on BigLinux. Kokoro was checked with the real koko: sound, a completed History entry, and valid float audio.