affbf6a9af6b7f4f249dda57d2c884dc84c61e94
2
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> |
||
|
Sergey SilaevandGitHub
|
582d494633 |
chore(build): uv tooling, a Python 3.11 floor, and verified model downloads (epic #701, phases 2.5 & 5) (#713)
* chore(build): migrate justfile tooling to uv All recipes now run through 'uv run' against .venv, with a shared DEV_EXTRAS variable (dev + vad) because uv syncs the environment exactly and would otherwise uninstall the dev tools between recipes. 'deps' installs dev+vad; new 'deps-all' covers every extra. 'build' switches to 'uv build'; 'version' uses --no-sync. Bump [tool.uv] required-version to >=0.12,<0.13 for uv 0.12.x, and raise the mypy python_version target to 3.12: mypy 1.19 in the lock cannot target 3.9 and numpy 2.x stubs need a 3.12 target to parse (mypy is not run in CI or pre-commit; this only affects 'just typecheck'). * fix(tests): isolate test runs from real XDG state; fix Arch model-path crash Local test runs failed en masse (185 failures) while CI stayed green: tests picked up the developer's real ~/.local/share/vocalinux models via the default XDG paths and switched IBus/keyboard code paths through the desktop session env (XDG_CURRENT_DESKTOP/ XDG_SESSION_TYPE on GNOME/Wayland, issue #478 branches). conftest now points XDG base dirs at a throwaway session-wide directory and clears session variables at import time, mirroring a fresh CI runner (paths like MODELS_DIR freeze at first import during collection, so fixture-time isolation is too late). Also guard the Arch branch of _get_system_model_paths(): it removed /usr/local/share/vocalinux/models unconditionally and crashed with ValueError on Arch hosts when a custom XDG_DATA_DIRS never added that path. Add regression tests, and make the lock-file test assert the actual $XDG_DATA_HOME/vocalinux contract instead of the hardcoded ~/.local/share default. * fix(tests): keep the IBus engine double a plain class Python 3.14's unittest.mock re-instantiates a user subclass when it builds child mocks (klass(name=...)). VocalinuxEngine inherits from whatever IBus.Engine resolves to, and the test module bound that to the MagicMock class itself, so every engine instantiation raised TypeError: VocalinuxEngine.__init__() got an unexpected keyword argument 'name' on 3.14 while passing on 3.13. Replace it with a plain stand-in class that still hands out per-instance MagicMocks for the IBus API surface the engine touches. The same pattern remains in test_ibus_engine.py, test_ibus_engine_utils.py, test_ibus_active_engine_guard.py and test_ibus_xkb_wayland.py; those never instantiate the engine, so they are unaffected for now. * fix(install): build the venv from the system Python install.sh looked up a bare `python3` from PATH. Started from an activated virtualenv — a shell left in uv's .venv after `just deps` — it built venv/ from that interpreter, where distro PyGObject is invisible even with --system-site-packages, and the run aborted with Distro PyGObject (python3-gi / python3-gobject) is not importable in the venv. A mismatched venv/ then survived every retry, because the installer reuses whatever venv it finds. - drop an inherited VIRTUAL_ENV (and its bin from PATH) before any interpreter lookup - select the interpreter explicitly: $SYSTEM_PYTHON (default /usr/bin/python3) or PATH python3, preferring one that can import gi, and create the venv with it instead of `python3` - recreate venv/ when it was built by a different interpreter - name the interpreter in the PyGObject guard's error Documents the two environments (uv's .venv for tooling, venv/ for the installed app) in CONTRIBUTING, the justfile header and AGENTS.md, and adds the mismatch symptom to the INSTALL troubleshooting section. * chore(python): raise the supported floor to 3.11 requires-python promised 3.9 while the CI matrix ran 3.9/3.10/3.11/3.13, the classifiers advertised up to 3.14, mypy targeted 3.12 and install.sh enforced 3.9 — five different answers to "which Python does this support". Settle on 3.11 everywhere: pyproject (requires-python, classifiers, black targets), install.sh's MIN_VERSION, check-system-deps.sh, the distro package map, the CI matrix (now 3.11-3.14, matching the classifiers), the release and nightly workflows, and the docs. mypy now targets the floor instead of 3.12, so type checking covers what we promise; numpy 2.x ships stubs in 3.12 syntax that a 3.11 target cannot parse, so its imports are skipped. (mypy has 71 pre-existing errors either way; no CI job gates on it.) uv.lock and the requirements/* exports are regenerated — dropping 3.9/3.10 removes most of the version-conditional resolution branches. tests/test_python_version_policy.py pins the invariant: install.sh's floor, requires-python, the classifiers and the CI matrix must agree. * ci: run lint and tests from uv.lock The lock file and the hash-pinned exports were not enforced anywhere: CI still built a venv by hand and installed unpinned `PyGObject` plus `-e ".[dev,vad]"`, so the tool versions it checked against had nothing to do with the ones `just lint` uses locally. Both Python jobs now use astral-sh/setup-uv and run through `uv sync --locked` / `uv run --locked`, which also fails the build when uv.lock drifts from pyproject.toml. Partially covers phase 5 of #701 (tests install from the lock, lock freshness gate); trusted publishing, artifact checksums and action SHA pinning are still open there. * ci: lint from a dedicated dependency group The lint job failed on this branch: `uv run --extra dev` syncs the whole project, so it tried to build pyaudio from source on a runner that installs no system dependencies. src/pyaudio/device_api.c:9:10: fatal error: portaudio.h: No such file or directory Move black/isort/flake8 out of the `dev` extra into a `lint` dependency group. CI runs `uv run --locked --only-group lint`, which installs the three pure-Python tools and skips the project entirely — no portaudio or girepository headers needed. `just deps` and DEV_EXTRAS request the group alongside the dev extras, so the local environment still carries test and lint tooling at once and `just lint` matches CI version for version. pre-commit is unaffected: its hooks pin their own tool versions upstream. * fix(install): never configure an engine the venv cannot import vosk became an optional extra in #705, but the installer still writes `"engine": "vosk"` into config.json on the whisper.cpp fallback path even when installing the extra only produced a warning. config.json also survives reinstalls — every writer guards on "file does not exist" — so an engine picked by an earlier attempt outlives the venv that supported it. Either way Vocalinux dies at startup: ModuleNotFoundError: No module named 'vosk' - the fallback now writes the vosk config only when `import vosk` actually succeeds, and otherwise leaves the engine configuration alone - a new verify_configured_engine() runs near the end of the install: if config.json selects vosk and the venv cannot import it, it says so and prints both fixes instead of leaving a broken install behind * feat(models): verify every model download against a pinned digest Pins a sha256/md5 per downloadable speech model in src/vocalinux/utils/model_checksums.txt and verifies against it in both install.sh and the runtime downloaders, before the file is moved into place. Verification fails closed. Covers phase 2.5 of #701; see the PR for details. * docs(agents): rebase AGENTS.md onto the #727 rewrite * fix(install): hash models that are already on disk instead of trusting them * refactor(models): one download path per engine, and fix Whisper's * docs(install): drop distros that cannot meet the 3.11 floor * docs: retire the distros and Python versions the 3.11 floor rules out * fix(install): abort instead of warning when Python is below the floor * fix(install): fall back to the system Python, not the one on PATH * fix(install): match an existing venv by installation, not by X.Y * fix(install): repair a config naming an engine the venv cannot import * fix(models): pin VOSK by the sha256 of bytes we fetched, not upstream's md5 * fix(justfile): request the dev set in every recipe that syncs * test(version): scan the docs and website trees for a stale Python floor * docs: correct the stale model-checksum and distro claims * fix(models): stamp the VOSK tree the app extracts, as the installer does * fix(install): replace an unverified VOSK model only once its successor is ready * docs(install): align the Fedora floor with the rest of the tree * test(install): drop an f-string with no placeholders * test(models): stub the downloader probe the VOSK installer test relies on * docs: drop the long dashes from two model-checksum comments * fix(install): gate on the interpreter, not on the distro release label * fix(justfile): export the lint group into requirements/dev.txt * fix(models): unpack the VOSK tree into staging and stamp it before the swap * fix(whisper): adopt a checkpoint an earlier release stored under the catalog name * docs(settings): state what a stray Whisper copy really means for deletion * fix(models): hash a whisper.cpp model found on disk before it reaches ctypes * docs(contributing): stop telling readers to uninstall their linters * fix(install): stop warning off distros the project documents as supported * docs(distros): file the supported derivatives under supported |