8f4aeca2f10f8f757bb697e593995ed09d272da6
4
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
8f4aeca2f1 |
feat(asr): add Parakeet TDT 0.6B engine via sherpa-onnx (#802)
* feat(asr): add Parakeet engine via sherpa-onnx Adds a "parakeet" engine backed by sherpa-onnx, running NVIDIA Parakeet TDT 0.6B int8 models on CPU: v3 (25 European languages, auto-detected) and v2 (English). Output is punctuated and capitalized. On an Intel Core Ultra 7 155H it decodes at about 12x realtime on CPU (RTF 0.08), roughly 8x faster than whisper.cpp large-v3-turbo over the same clips with its Vulkan GPU backend (RTF 0.59). Being a transducer it also has no 30s window to pad, so a short push-to-talk utterance returns in ~0.3s instead of ~2.4s. Follows the existing engine layout: utils/parakeet_model_info.py holds the catalog and the get_model_path / is_model_downloaded / list_downloaded_models / delete_model helpers, while _init_parakeet / _transcribe_with_parakeet / _download_parakeet_model follow the same shape as their whisper.cpp counterparts and reuse _stream_model_download for progress and cancellation. Registered in the engine dispatch, the settings engine/model/language pickers, the model info card, the download prompts, the unused-downloads list and install.sh. Models download to <models_dir>/parakeet/<model>/ via temp files renamed on success; is_model_downloaded requires all four files, so an interrupted download is retried rather than loaded. Requires the new "parakeet" extra (sherpa-onnx). * fix(install): write only the engine key for parakeet The parakeet branch duplicated a full default config.json. ConfigManager deep-merges DEFAULT_CONFIG on load, so every other key was redundant: a config holding only "engine": "parakeet" still resolves parakeet_model_size to v3-european along with the VAD, shortcut and text-injection defaults. * fix(install): keep push-to-talk defaults and switch existing configs Writing only the engine key left config.json without a shortcuts section, which _migrate_shortcuts_config reads as a pre-push-to-talk upgrade and pins back to ctrl+ctrl/toggle - contradicting the installer's own "Hold Right Alt while you speak" message. Seed that section too. Point the config at parakeet with set_configured_engine rather than only on first create, so --engine=parakeet also switches a machine that already has a config.json, and give print_welcome_message a parakeet arm so the summary is not blank. The interactive picker overwrites SELECTED_ENGINE, so document the flag combination that reaches this branch. * fix(asr): pin Parakeet bundles and report one progress bar Pin each bundle to a Hugging Face commit instead of resolve/main, so a file replaced upstream cannot change what an install downloads. This is the same approach whispercpp_model_info takes with its pinned revision. Remap each file's progress into its slice of the bundle. The four files were reported through the single-file downloader independently, so the fraction ran to 1.0 and restarted four times; it now advances once and names the file being fetched. * fix(asr): reject a truncated Parakeet download _stream_model_download only rejects an empty body, so a truncated but nonempty response was renamed into the model directory and then reported as installed, because is_model_downloaded checks existence. Record the byte size of each file at the pinned revision and check it before the rename, the same order whisper.cpp verifies its digest in. A short file is a truncated transfer rather than an upstream change, since the revision is pinned, so it raises and the temp file is removed. * style(asr): annotate the Parakeet model catalog mypy inferred the catalog values as object, so reading file_sizes out of an entry raised attr-defined and no-any-return under `just typecheck`. * fix(asr): pin Parakeet digests in model_checksums.txt AGENTS.md requires every speech model to be verified against a pinned digest before install, failing closed on an unpinned one. Parakeet was checking only the byte size. Teach scripts/generate-model-checksums.py to emit Parakeet entries: the ONNX files are LFS so Hugging Face publishes their sha256, and tokens.txt is hashed from the downloaded bytes the way VOSK archives are. Several bundles ship the same file names, so a key carries the model name. _download_parakeet_model now calls verify_model_file before the rename and catches ChecksumError alongside the other download failures, matching the whisper.cpp downloader. * fix(asr): verify on-disk Parakeet models before load Existing bundle files were trusted by path alone, so a leftover or copied-in tree could reach sherpa-onnx unchecked. Hash each file against its pin on download skip and on init, and cover the Parakeet pins in CI. * fix(ui): hide Parakeet language picker that ignores the setting Parakeet language coverage is the model choice, not a catalog language. Hide the picker and persist auto so Settings does not write a lying value. * fix(asr): normalize Parakeet language to auto at startup CLI --language and stale saved languages still reached SpeechRecognitionManager even though Parakeet ignores catalog language. Force auto on the shared helper used by main and the manager so stored settings stay honest. * fix(ui): restore language after leaving Parakeet Remember the prior Whisper/cpp language across a Parakeet visit so switching back does not leave recognition stuck on auto. Keep Parakeet forced to auto and the picker hidden. Also satisfy black on the startup normalizer path. * fix(ui): keep language memory across Vosk fallback Vosk's unsupported-language en-us coercion was writing into _last_non_parakeet_language, so Whisper→Parakeet→Vosk→Whisper restored en-us instead of the user's Greek (or other) preference. Remember the pre-coercion catalog language and prefer it when leaving Vosk. * fix(ui): keep language memory when entering Parakeet from Vosk Vosk's coerced en-us fallback was still written into _last_non_parakeet_language on Parakeet entry, so Vosk→Parakeet→Whisper restored en-us instead of the pre-coercion preference (Greek or auto). Share one coerced-fallback predicate across entry, leave-Vosk, and language sync, and cover those paths with regression tests. * fix(tests): set initializing flags false on language-change stubs Plain Mock attributes are truthy and were short-circuiting the handler. --------- Co-authored-by: Alex Thunder <AlexThunder@gmail.com> |
||
|
Jatin K MalikandGitHub
|
40bc31bebf |
fix(settings): stop freezing when applying an already-downloaded model (#790)
* fix(settings): apply already-downloaded models off the GTK main loop Changing Model Size, Specialization, or Language froze Settings when the model was already on disk, because reconfigure ran in the changed handler. Move that apply onto a worker thread (same shape as the download path) and keep the apply-guard held until GLib.idle_add clears it so a second pick cannot start a parallel load. Fixes #789. * fix(settings): guard Test/apply while downloaded-model worker runs Block apply_settings and Test Dictation while _applying_settings is held so a concurrent Test click cannot race the shared engine. Annotate the new worker/helpers per AGENTS.md and cover the no-op/block behavior. * fix(settings): resync pickers when a second pick lands mid-apply While an already-downloaded model worker was running, Model Size / Specialization / Language could still move. The apply guard blocked saving the new pick, then finish only cleared the flag, so the UI could show B with dictation still on A. On finish, if the selection no longer matches the saved config, put the pickers back. Idle resyncs from workers now skip a destroyed dialog the same way update check does. * fix(settings): always resync pickers after model apply finishes A second Model Size pick on whisper.cpp could move the size combo while the apply guard skipped rebuilding specialization, so selected settings still looked like the saved variant and finish skipped the resync. Always put the pickers back on the saved config when the worker ends, and restore language from that config too. * test(settings): prove mid-apply picker handlers round-trip on finish Call the real size, specialization, and language handlers while the apply guard is held, and run finish through the real resync path so whisper.cpp size and language cannot silently stick on the second pick. |
||
|
|
6b14338d1e |
feat(tray): offer the recommended model from the missing-model notification (#687)
* feat(tray): offer the recommended model from the missing-model notification Dictation without a model on disk showed a dead-end notification telling the user to open Settings, find the model picker and start the download by hand, at the moment they wanted to talk. There was no Download button anywhere: downloads only happen as a side effect of changing a selection. The notification now carries a "Download <model> (<size>)" button for the model the picker already stars for this machine, so recommendations stay in one place: get_recommended_model() for whisper.cpp, including the English-only variant when an English language is selected, and the RAM based recommendations for VOSK and Whisper. Clicking it downloads in the background, reports progress in the notification, and saves the model for that engine only after it loaded. Notification servers that do not advertise the actions capability, and systems without libnotify, still get the plain text they got before. Dictation is not started for the user afterwards: a download can take several minutes, and grabbing the microphone unannounced would surprise someone who has long since moved on. * feat(tray): actually fetch the offered model, and only claim it when loaded Review follow-ups on #687. reconfigure() only re-initializes when engine, model or language change, and force_download is read during that re-init. First-run defaults are whisper_cpp + tiny, and tiny is what get_recommended_model() returns on a CPU-only machine under 16GB, so the click was a no-op that logged "Applied VAD/silence timeout changes.", saved the model, said "Speech model ready", and left the next shortcut showing the same missing-model notification. Add force_reinit to reconfigure() and pass it here, then gate the save and the success notification on speech_engine.model_ready instead of on reconfigure() not raising. The "starting..." notification is now created in _download_recommended_model, which runs on the main loop where libnotify dispatches the action; only the handle travels to the worker, which already routes update/close/notify through _idle_once. The offer body no longer promises a button the server cannot draw: with no action it points at Settings, like the notification it replaced. Tests cover the case that breaks the feature — the recommendation equal to the configured model — behaviourally, through a stand-in engine that gates re-init the way reconfigure() does, plus the real reconfigure() contract in both directions. They also stop touching sys.modules at collection time, which is the desync #686 hit, and black is clean. * test(notifications): cover the libnotify/notify-send boundary directly Every test that touches notifications so far patches the whole module away, so the module itself had no coverage: the fallback to notify-send when the server has no action support or libnotify fails to initialize, the handle kept alive in _pending until the user answers, and update() refusing a fallback handle. These drive it with a fake Notify installed through the module's own lazy cache rather than by patching gi.repository, which is shared with every other test in the process. * fix(download): allow only one model download at a time across tray and Settings The tray's download worker and the Settings dialog's both drive the same engine, which has a single progress callback slot and a single configuration. Started together, the later one overwrote the earlier one's progress callback, the earlier one's finally block then cleared it mid-download, and whichever finished last became the persisted model. Give the engine a non-blocking claim. Both Settings entry points and the tray take it before starting a download and release it when the worker ends; whoever loses the race is told a download is already running and the Settings picker goes back to the saved model. --------- Co-authored-by: Kacper Paczos <23406555+kacperpaczos@users.noreply.github.com> |
||
|
Kacper PaczosandGitHub
|
a4bd398945 |
fix(settings): save the model only once the engine loaded it (#685)
* fix(settings): save the model only once the engine loaded it Applying settings wrote the config first and reconfigured the engine afterwards, and reconfiguring is what downloads a missing model. A download that failed, was cancelled, or was interrupted therefore left the config naming a model that never reached the disk: the next start reported "No Speech Model" while the previously working model sat on disk, deselected. Reconfigure first and persist only on success, so a failed switch keeps the settings that are actually running. On failure the pickers are put back on the saved model. They otherwise kept showing the rejected one, which also blocked a retry, since re-selecting the same entry emits no "changed" signal. Selecting an engine now applies as well. Every other control on the page applies on change, but the engine only repainted the pickers: the config and the recognizer stayed on the previous engine until another control was touched, and closing the dialog dropped the change without a word. Remote API is deferred until its server URL is filled in, since applying it empty can only fail validation. * fix(settings): restore the engine picker after a failed engine switch Fixes the known limitation called out in the previous commit (#692): when applying a newly selected engine failed, the resync restored the model pickers but left the engine combo displaying the engine that was never applied, so the page described a state that was not running. Set the combo back to the saved engine while holding the apply suppression flag, so _on_engine_changed() repaints the pickers without cascading into another apply or download prompt. * fix(settings): resync the engine picker on every path that never applied Review follow-ups on #685. #692 was not actually closed. The resync lived only in the download thread's except block, and _apply_settings_internal() reports failure by returning False as well as by raising, so the thread fell through to set_complete(True, "") and reported a switch that never happened. Act on the return value instead of depending on #690 to make that except live, and add the same handling to apply_settings()'s modal path, whose except never resynced at all. Add _resync_engine_ui_if_unapplied() as the backstop the review asked for: after the modal closes, and when the dialog closes, compare the engine on screen against the engine in the config and restore the picker when they disagree. That also covers Remote API left without a server URL, where applying stays deliberately deferred. _on_engine_changed now checks _initializing / _applying_settings itself, like every other control on this page, so the combo move made by a resync repaints without rewriting the user's language or cascading into another apply. Tests no longer touch sys.modules at collection time: the reimport moved into a module-scoped fixture that also restores the attribute on vocalinux.ui, which is what desynced test_model_deletion on 3.9/3.10. New cases cover the modal path that #692 is about — needs_download true with the apply returning False, and with it raising "Download cancelled" — plus the close hook. The stub now uses RecognitionState.IDLE rather than the string "idle", so tests stop taking the stop_recognition branch and sleeping 0.5s. |