python-voiceio 1.15.1__tar.gz → 1.16.0__tar.gz

{python_voiceio-1.15.1/python_voiceio.egg-info → python_voiceio-1.16.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 1.15.1
+Version: 1.16.0
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT
@@ -146,6 +146,10 @@ Everything lives in `~/.config/voiceio/config.toml`; [config.example.toml](confi
 
 Open an issue before a large PR. [CONTRIBUTING.md](CONTRIBUTING.md) covers the architecture, conventions and the reasoning behind the model choice. Install reports and ideas are welcome on the [feedback board](https://voiceio.dev/#feedback).
 
+## How it's built
+
+I write voiceio with AI coding assistants, mostly Claude Code. I decide what goes in and I use it every day, which is how most of its bugs get found.
+
 ## Credits
 
 - Speech recognition: [faster-whisper](https://github.com/SYSTRAN/faster-whisper) and [CTranslate2](https://github.com/OpenNMT/CTranslate2), with OpenAI's [Whisper](https://github.com/openai/whisper) models; [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) with NVIDIA's [Parakeet-TDT](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2). The in-browser demo runs [Moonshine](https://github.com/usefulsensors/moonshine) through [transformers.js](https://github.com/huggingface/transformers.js).

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/README.md

@@ -93,6 +93,10 @@ Everything lives in `~/.config/voiceio/config.toml`; [config.example.toml](confi
 
 Open an issue before a large PR. [CONTRIBUTING.md](CONTRIBUTING.md) covers the architecture, conventions and the reasoning behind the model choice. Install reports and ideas are welcome on the [feedback board](https://voiceio.dev/#feedback).
 
+## How it's built
+
+I write voiceio with AI coding assistants, mostly Claude Code. I decide what goes in and I use it every day, which is how most of its bugs get found.
+
 ## Credits
 
 - Speech recognition: [faster-whisper](https://github.com/SYSTRAN/faster-whisper) and [CTranslate2](https://github.com/OpenNMT/CTranslate2), with OpenAI's [Whisper](https://github.com/openai/whisper) models; [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) with NVIDIA's [Parakeet-TDT](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2). The in-browser demo runs [Moonshine](https://github.com/usefulsensors/moonshine) through [transformers.js](https://github.com/huggingface/transformers.js).

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "python-voiceio"
-version = "1.15.1"
+version = "1.16.0"
 description = "Voice dictation for Linux. Speak → text, locally, instantly."
 readme = "README.md"
 license = "MIT"

{python_voiceio-1.15.1 → python_voiceio-1.16.0/python_voiceio.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 1.15.1
+Version: 1.16.0
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT
@@ -146,6 +146,10 @@ Everything lives in `~/.config/voiceio/config.toml`; [config.example.toml](confi
 
 Open an issue before a large PR. [CONTRIBUTING.md](CONTRIBUTING.md) covers the architecture, conventions and the reasoning behind the model choice. Install reports and ideas are welcome on the [feedback board](https://voiceio.dev/#feedback).
 
+## How it's built
+
+I write voiceio with AI coding assistants, mostly Claude Code. I decide what goes in and I use it every day, which is how most of its bugs get found.
+
 ## Credits
 
 - Speech recognition: [faster-whisper](https://github.com/SYSTRAN/faster-whisper) and [CTranslate2](https://github.com/OpenNMT/CTranslate2), with OpenAI's [Whisper](https://github.com/openai/whisper) models; [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) with NVIDIA's [Parakeet-TDT](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2). The in-browser demo runs [Moonshine](https://github.com/usefulsensors/moonshine) through [transformers.js](https://github.com/huggingface/transformers.js).

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/python_voiceio.egg-info/SOURCES.txt

@@ -29,6 +29,7 @@ tests/test_fallback.py
 tests/test_fcitx_bridge.py
 tests/test_feedback_stream.py
 tests/test_focus.py
+tests/test_gpu_fallback.py
 tests/test_health.py
 tests/test_hints.py
 tests/test_history.py

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/tests/test_feedback_stream.py

@@ -1,11 +1,11 @@
 """The persistent cue stream: buffer size and cue timing."""
 from __future__ import annotations
 
-import pytest
+from types import SimpleNamespace
 
-pytest.importorskip("sounddevice")
+import pytest
 
-from voiceio import feedback  # noqa: E402
+from voiceio import feedback
 
 
 class FakeStream:
@@ -24,7 +24,7 @@ class FakeStream:
 @pytest.fixture
 def fake_stream(monkeypatch):
     FakeStream.opened = []
-    monkeypatch.setattr(feedback.sd, "OutputStream", FakeStream)
+    monkeypatch.setattr(feedback, "sd", SimpleNamespace(OutputStream=FakeStream))
     monkeypatch.setattr(feedback, "_stream", None)
     monkeypatch.setattr(feedback, "_sounds", {})
     return FakeStream

python_voiceio-1.16.0/tests/test_gpu_fallback.py

@@ -0,0 +1,80 @@
+"""device = "auto" with a GPU whose CUDA libraries are missing: CPU, not a crash loop."""
+from __future__ import annotations
+
+import sys
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+
+worker = pytest.importorskip("voiceio.core.decoders.faster_whisper.worker")
+
+ARGS = {"model": "small", "compute_type": "int8", "language": "en"}
+
+
+def _loader(fail_on: set[str]):
+    devices = []
+
+    def load(args, device):
+        devices.append(device)
+        if device in fail_on:
+            raise RuntimeError("Library libcublas.so.12 is not found or cannot be loaded")
+        return SimpleNamespace(device=device)
+    return load, devices
+
+
+def test_auto_falls_back_to_cpu_when_cuda_fails_at_first_encode(capsys):
+    load, devices = _loader({"auto"})
+    with patch.object(worker, "_load_and_warm", load):
+        model, why = worker.load_ready({**ARGS, "device": "auto"})
+    assert model.device == "cpu" and devices == ["auto", "cpu"]
+    assert "libcublas.so.12" in why
+    assert "decoding on CPU" in capsys.readouterr().err
+
+
+def test_explicit_cuda_still_fails_loudly():
+    load, _ = _loader({"cuda"})
+    with patch.object(worker, "_load_and_warm", load), pytest.raises(RuntimeError):
+        worker.load_ready({**ARGS, "device": "cuda"})
+
+
+def test_working_device_reports_no_fallback():
+    load, devices = _loader(set())
+    with patch.object(worker, "_load_and_warm", load):
+        assert worker.load_ready({**ARGS, "device": "auto"})[1] == ""
+    assert devices == ["auto"]
+
+
+def test_doctor_names_missing_cuda_libs(monkeypatch):
+    from voiceio import health
+    from voiceio.config import Config
+    cfg = Config()
+    cfg.model.name, cfg.model.device = "small", "auto"
+    monkeypatch.setattr("voiceio.config.load", lambda: cfg)
+    monkeypatch.setitem(sys.modules, "ctranslate2",
+                        SimpleNamespace(get_cuda_device_count=lambda: 1))
+    real = __import__("ctypes").CDLL
+
+    def cdll(name, *a, **k):
+        if name == "libcublas.so.12":
+            raise OSError(name)
+        return real("libc.so.6")
+    monkeypatch.setattr("ctypes.CDLL", cdll)
+    msg = health._gpu_libs_warning()
+    assert "libcublas.so.12" in msg and "libcudnn" not in msg and 'model.device = "cpu"' in msg
+    cfg.model.device = "cpu"
+    assert health._gpu_libs_warning() == ""
+
+
+def test_daemon_accepts_the_fallback_handshake_and_logs_it(caplog):
+    from unittest.mock import MagicMock
+
+    from voiceio.config import ModelConfig
+    from voiceio.core.decoders.faster_whisper import Transcriber
+    proc = MagicMock()
+    proc.stdout.readline.return_value = "READY cpu-fallback: libcublas.so.12 missing\n"
+    proc.poll.return_value = None
+    with patch("voiceio.core.decoders.faster_whisper.subprocess.Popen", return_value=proc), \
+         caplog.at_level("WARNING"):
+        Transcriber(ModelConfig())
+    assert "decoding on CPU instead: libcublas.so.12 missing" in caplog.text

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/tests/test_health.py

@@ -149,3 +149,16 @@ def test_check_ibus_asks_about_voiceio_only_when_the_engine_is_silent():
             alive.reset_mock()
             assert health._check_ibus().voiceio_running is True
             alive.assert_not_called()
+
+
+def test_kde_wayland_without_ibus_input_method_warns(tmp_path):
+    from voiceio.health import _kde_input_method_warning
+    kde = Platform(os="linux", display_server="wayland", desktop="kde")
+    rc = tmp_path / "kwinrc"
+    rc.write_text("[Wayland]\nInputMethod[$e]=/usr/share/applications/com.github.maliit.keyboard.desktop\n")
+    assert "Virtual Keyboard" in _kde_input_method_warning(kde, rc)
+    rc.write_text("[Wayland]\nInputMethod[$e]=/usr/share/applications/org.freedesktop.IBus.Panel.Wayland.Gtk3.desktop\n")
+    assert _kde_input_method_warning(kde, rc) == ""
+    assert "Virtual Keyboard" in _kde_input_method_warning(kde, tmp_path / "missing")
+    gnome = Platform(os="linux", display_server="wayland", desktop="gnome")
+    assert _kde_input_method_warning(gnome, tmp_path / "missing") == ""

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/tests/test_idle_audio.py

@@ -3,17 +3,16 @@ from __future__ import annotations
 
 import threading
 import time
+from types import SimpleNamespace
 from unittest.mock import MagicMock, patch
 
 import numpy as np
 import pytest
 
-pytest.importorskip("sounddevice")
-
-from voiceio import feedback, idle_audio  # noqa: E402
-from voiceio.config import AudioConfig, Config  # noqa: E402
-from voiceio.idle_audio import IdleAudio  # noqa: E402
-from voiceio.recorder import AudioRecorder  # noqa: E402
+from voiceio import feedback, idle_audio
+from voiceio.config import AudioConfig, Config
+from voiceio.idle_audio import IdleAudio
+from voiceio.recorder import AudioRecorder
 
 
 class FakeInput:
@@ -50,8 +49,9 @@ class FakeOutput(FakeInput):
 @pytest.fixture
 def streams(monkeypatch):
     FakeInput.opened = FakeOutput.opened = 0
-    monkeypatch.setattr("voiceio.recorder.sd.InputStream", FakeInput)
-    monkeypatch.setattr(feedback.sd, "OutputStream", FakeOutput)
+    # Whole fakes: touching the real module would initialize PortAudio.
+    monkeypatch.setattr("voiceio.recorder.sd", SimpleNamespace(InputStream=FakeInput))
+    monkeypatch.setattr(feedback, "sd", SimpleNamespace(OutputStream=FakeOutput))
     monkeypatch.setattr(feedback, "_stream", None)
 
 

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/tests/test_postprocess.py

@@ -88,6 +88,20 @@ class TestStripDisfluencies:
         assert strip_disfluencies("uh-huh right") == "right"
         assert strip_disfluencies("uh huh yes") == "yes"
 
+    def test_mm_hmm_goes_whole(self):
+        # Parakeet ends about 1 clip in 4 with an invented "Mm-hmm."; cutting
+        # only its "hmm" typed a stray "Mm-." after real dictation.
+        assert strip_disfluencies("I sent it. Mm-hmm.") == "I sent it."
+        assert strip_disfluencies("Are you sure? Mm-hmm.") == "Are you sure?"
+        assert strip_disfluencies("Mm-hmm.") == ""
+        assert strip_disfluencies("Fine. Mm-hmm. So it works.") == "Fine. So it works."
+        assert strip_disfluencies("mmhmm that works") == "that works"
+
+    def test_fillers_inside_hyphenated_words_survive(self):
+        assert strip_disfluencies("a hmm-hmm sound") == "a hmm-hmm sound"
+        assert strip_disfluencies("she said uh-uh twice") == "she said uh-uh twice"
+        assert strip_disfluencies("the 5 mm drill") == "the 5 mm drill"
+
     def test_preserves_emphatic_short_repeat(self):
         # Short repeats are emphasis, not a re-decode artifact — keep them.
         assert strip_disfluencies("No. No.") == "No. No."

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/tests/test_report.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 import io
 import json
 import sys
+from pathlib import Path
 
 import pytest
 
@@ -96,12 +97,12 @@ def test_report_contains_only_backend_and_model_names(identity):
 def test_report_lists_failures_and_warnings():
     health = _health(typer_backends=[BackendStatus("ibus", False, "daemon down")],
                      cli_in_path=False, tray_ok=False, tray_reason="no AppIndicator")
-    text = report.build_report(health, missing_deps=["portaudio"], perm_issues=2, cfg=Config())
+    text = report.build_report(health, missing_deps=["portaudio"], perm_issues=[Path(FAKE_HOME) / ".config/voiceio/config.toml"], cfg=Config())
     assert "typer none" in text
     assert "typer ibus: daemon down" in text
     assert "missing system deps" in text
     assert "cli not in PATH" in text and "tray: no AppIndicator" in text
-    assert "insecure permissions on 2 file(s)" in text
+    assert "insecure permissions on ~/.config/voiceio/config.toml" in text
 
 
 def test_silent_ibus_engine_is_expected_while_voiceio_is_stopped():
@@ -109,9 +110,9 @@ def test_silent_ibus_engine_is_expected_while_voiceio_is_stopped():
     ib = IBusStatus(component_installed=True, daemon_running=True, env_persisted=True)
     stopped = report.build_report(_health(ibus_checks=IBusStatus(**{**vars(ib), "voiceio_running": False})),
                                   cfg=Config())
-    assert "engine ping" not in stopped and "voiceio not running" in stopped
+    assert "engine ping" not in stopped and "voiceio was not running when this report was made" in stopped
     running = report.build_report(_health(ibus_checks=ib), cfg=Config())
-    assert "ibus engine ping" in running and "voiceio not running" not in running
+    assert "ibus engine ping" in running and "not running" not in running
 
 
 def test_comment_is_one_bounded_line():
@@ -121,6 +122,34 @@ def test_comment_is_one_bounded_line():
     assert len(line) <= len("comment: ") + report.MAX_COMMENT
 
 
+def test_long_comment_ends_with_an_ellipsis_at_a_word():
+    text = report.build_report(_health(), comment="word " * 400, cfg=Config())
+    line = next(ln for ln in text.splitlines() if ln.startswith("comment:"))
+    assert line.endswith("word…") and len(line) <= len("comment: ") + report.MAX_COMMENT
+    short = report.build_report(_health(), comment="fine " * 100, cfg=Config())
+    assert "…" not in short
+
+
+def test_works_line_only_for_a_known_answer():
+    assert "works: no" in report.build_report(_health(), works="no", cfg=Config()).splitlines()
+    assert "works:" not in report.build_report(_health(), works="maybe", cfg=Config())
+
+
+def test_many_insecure_paths_are_counted(identity):
+    paths = [Path(FAKE_HOME) / f"f{i}" for i in range(7)]
+    text = report.build_report(_health(), perm_issues=paths, cfg=Config())
+    assert "~/f0, ~/f1, ~/f2, ~/f3, ~/f4 and 2 more" in text
+    _assert_private(text)
+
+
+def test_ask_details_maps_answers(monkeypatch):
+    answers = iter(["  it typed twice  ", "Y"])
+    monkeypatch.setattr(report, "_ask", lambda prompt: next(answers))
+    assert report.ask_details() == ("it typed twice", "yes")
+    answers = iter(["", "whatever"])
+    assert report.ask_details() == ("", "")
+
+
 def test_comment_drops_terminal_escapes():
     # A real 1.15.0 report arrived with "comment: \x1bv" (Alt+V at the prompt).
     text = report.build_report(_health(), comment="\x1bv", cfg=Config())

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/tests/test_robustness.py

@@ -936,3 +936,20 @@ class TestZombieStreamRecovery:
         # any real signal → not a zombie
         rec._ring.append(np.full(rec._ring._max, 0.1, dtype=np.float32))
         assert rec.is_zombie() is False
+
+
+def test_app_imports_when_portaudio_cannot_initialize(tmp_path):
+    # ubuntu-26.04 runners: libportaudio2 pulls libpulse0, and with no pulse
+    # server `import sounddevice` raises; importing voiceio.app must not.
+    import subprocess
+    import sys
+    (tmp_path / "sounddevice.py").write_text(
+        "class PortAudioError(Exception): pass\n"
+        "raise PortAudioError('PulseAudio_Initialize: Can\\'t connect to server')\n")
+    code = ("import voiceio.app, voiceio.recorder as r\n"
+            "try:\n    r.sd.InputStream\nexcept Exception as e:\n"
+            "    print(type(e).__name__)\n")
+    out = subprocess.run([sys.executable, "-c", code], capture_output=True, text=True,
+                         env={**__import__("os").environ, "PYTHONPATH": str(tmp_path)})
+    assert out.returncode == 0, out.stderr
+    assert out.stdout.strip() == "PortAudioError"

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/tests/test_service.py

@@ -168,3 +168,29 @@ def test_reinstalling_the_service_refreshes_learning_units_that_are_on(monkeypat
     monkeypatch.setattr(service, "refresh_learn_timer", lambda: calls.append("refresh"))
     monkeypatch.setattr(service.subprocess, "run", lambda *a, **k: None)
     assert service.install_service() and calls == ["refresh"]
+
+
+def test_slow_daemon_reload_still_enables_the_service(tmp_path, monkeypatch):
+    # Mint report: daemon-reload hit its 5s timeout and setup never ran enable.
+    import subprocess
+    from voiceio import service
+    monkeypatch.setattr(service, "has_systemd", lambda: True)
+    monkeypatch.setattr(service, "_find_voiceio_bin", lambda: "voiceio")
+    monkeypatch.setattr(service, "SERVICE_DIR", tmp_path)
+    monkeypatch.setattr(service, "SERVICE_PATH", tmp_path / service.SERVICE_NAME)
+    for name in ("install_symlinks", "uninstall_correct_timer", "refresh_learn_timer"):
+        monkeypatch.setattr(service, name, lambda: None)
+    ran = []
+
+    def run(cmd, **kw):
+        ran.append(cmd[2])
+        assert kw["timeout"] == service.SYSTEMCTL_TIMEOUT
+        if cmd[2] == "daemon-reload":
+            raise subprocess.TimeoutExpired(cmd, kw["timeout"])
+        return subprocess.CompletedProcess(cmd, 1 if cmd[2] == enable_fails else 0)
+    monkeypatch.setattr(service.subprocess, "run", run)
+    enable_fails = None
+    assert service.install_service()
+    assert ran == ["daemon-reload", "disable", "enable"]
+    enable_fails = "enable"
+    assert not service.install_service()

python_voiceio-1.16.0/voiceio/__init__.py

@@ -0,0 +1 @@
+__version__ = "1.16.0"

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/cli/doctor.py

@@ -87,7 +87,7 @@ def _cmd_doctor(args: argparse.Namespace) -> None:
         from voiceio.hints import hint
         hint("vocab_suggest", "Run 'voiceio vocab suggest' to find names it keeps mishearing")
         if args.report:
-            _share_report(args, report, missing_deps, len(perm_issues))
+            _share_report(args, report, missing_deps, [p for p, _, _ in perm_issues])
         sys.exit(0 if (report.all_ok and not perm_issues and not missing_deps) else 1)
 
     # Auto-fix mode
@@ -199,16 +199,19 @@ def _cmd_doctor(args: argparse.Namespace) -> None:
 
     if args.report:
         _share_report(args, report, plat.check_system_deps(),
-                      len(_config.check_permissions()))
+                      [p for p, _, _ in _config.check_permissions()])
     sys.exit(0 if report.all_ok else 1)
 
 
 def _share_report(args: argparse.Namespace, report, missing_deps: list[str],
-                  perm_issues: int) -> None:
-    from voiceio.report import build_report, share
+                  perm_issues: list) -> None:
+    from voiceio.report import _interactive, ask_details, build_report, share
+    comment, works = args.comment or "", ""
+    if not comment and not args.yes and _interactive():
+        comment, works = ask_details()
     try:
         text = build_report(report, missing_deps=missing_deps, perm_issues=perm_issues,
-                            comment=args.comment or "")
+                            comment=comment, works=works)
     except Exception as e:  # noqa: BLE001 — the report must not take doctor down
         print(f"\nCould not build the report ({type(e).__name__}); nothing was posted.")
         return

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/core/decoders/faster_whisper/__init__.py

@@ -134,12 +134,16 @@ class Transcriber:
 
         t0 = time.monotonic()
         ready = self._read_with_timeout(WORKER_START_TIMEOUT)
-        if ready is None or ready.strip() != "READY":
+        if ready is None or ready.split(" ", 1)[0].strip() != "READY":
             self._kill_worker()
             detail = "timed out" if ready is None else f"said {ready.strip()!r}"
             raise DecodeError(
                 f"worker failed to start ({detail} after {time.monotonic() - t0:.0f}s)"
             )
+        if fallback := ready.strip().partition("cpu-fallback: ")[2]:
+            log.warning("GPU unusable, decoding on CPU instead: %s "
+                        "(voiceio doctor shows the fix; model.device = \"cpu\" silences this)",
+                        fallback)
         elapsed = time.monotonic() - t0
         log.info("Model ready (%.1fs)", elapsed)
 

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/core/decoders/faster_whisper/worker.py

@@ -135,6 +135,39 @@ def decode(model: WhisperModel, audio: np.ndarray, kwargs: dict) -> list:
                                           kwargs.get("initial_prompt"))})
 
 
+def _load_and_warm(args: dict, device: str) -> WhisperModel:
+    model = load_model(
+        args["model"], device=device, compute_type=args["compute_type"],
+        cpu_threads=int(args.get("cpu_threads") or 0),
+    )
+    # Warmup: the first transcription is always slow. It also runs the first
+    # encode, where a CUDA build missing libcublas/libcudnn actually fails.
+    segs, _ = model.transcribe(np.zeros(16000, dtype=np.float32),
+                               language=args.get("language"), beam_size=1)
+    list(segs)
+    return model
+
+
+def load_ready(args: dict) -> tuple[WhisperModel, str]:
+    """The warmed-up model, and why it fell back to CPU ("" if it didn't).
+
+    device "auto" picks CUDA whenever a GPU is visible, even when the CUDA
+    libraries CTranslate2 needs are missing (a torch for another CUDA major
+    pulled in by the train extra): that failed every start and crash-looped
+    the service. On "auto" a GPU that can't run is a CPU decode instead;
+    an explicit "cuda" still fails loudly.
+    """
+    device = args["device"]
+    try:
+        return _load_and_warm(args, device), ""
+    except Exception as e:  # noqa: BLE001
+        if device != "auto":
+            raise
+        reason = " ".join(str(e).split())[:200] or type(e).__name__
+    print(f"voiceio: GPU unusable ({reason}); decoding on CPU", file=sys.stderr, flush=True)
+    return _load_and_warm(args, "cpu"), reason
+
+
 def main() -> None:
     # A hung decode is killed by the parent after a timeout; SIGUSR1 first
     # dumps every thread's stack to stderr (worker.log) so hangs are diagnosable.
@@ -142,15 +175,8 @@ def main() -> None:
     faulthandler.register(signal.SIGUSR1, all_threads=True)
 
     args = json.loads(sys.argv[1])
-    model = load_model(
-        args["model"], device=args["device"], compute_type=args["compute_type"],
-        cpu_threads=int(args.get("cpu_threads") or 0),
-    )
-
-    # Warmup: first transcription is always slow
-    segs, _ = model.transcribe(np.zeros(16000, dtype=np.float32), language=args.get("language"), beam_size=1)
-    list(segs)
-    print("READY", flush=True)
+    model, fallback = load_ready(args)
+    print(f"READY cpu-fallback: {fallback}" if fallback else "READY", flush=True)
 
     for line in sys.stdin:
         if line.strip() == "QUIT":

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/feedback.py

@@ -20,11 +20,9 @@ from pathlib import Path
 
 import numpy as np
 
-try:
-    import sounddevice as sd
-except ImportError as _e:
-    from voiceio.platform import desktop_dep_hint
-    raise ImportError(desktop_dep_hint("sounddevice")) from _e
+from voiceio.platform import LazySounddevice
+
+sd = LazySounddevice()
 
 log = logging.getLogger(__name__)
 

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/health.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 
 import logging
 from dataclasses import dataclass, field
+from pathlib import Path
 
 from voiceio import platform as plat
 from voiceio.hotkeys import chain as hotkey_chain
@@ -125,9 +126,66 @@ def check_health(p: plat.Platform | None = None) -> HealthReport:
                 "Grant it in System Settings → Privacy & Security → Accessibility."
             )
 
+    if gpu := _gpu_libs_warning():
+        report.warnings.append(gpu)
+    if kde := _kde_input_method_warning(report.platform):
+        report.warnings.append(kde)
+
     return report
 
 
+def _kde_input_method_warning(p: plat.Platform, kwinrc: Path | None = None) -> str:
+    """KWin on Wayland only hands text to the input method picked under
+    Virtual Keyboard; without IBus there, voiceio has no live preview."""
+    if not (p.desktop == "kde" and p.display_server == "wayland"):
+        return ""
+    import configparser
+    rc = configparser.ConfigParser(interpolation=None, strict=False)
+    try:
+        rc.read(kwinrc or Path.home() / ".config" / "kwinrc")
+    except configparser.Error:
+        return ""
+    # KWin writes it as "InputMethod[$e]" (a KConfig flag), so match the prefix.
+    section = rc["Wayland"] if rc.has_section("Wayland") else {}
+    if any(k.startswith("inputmethod") and "ibus" in v.lower() for k, v in section.items()):
+        return ""
+    return ("KDE isn't using IBus as its Wayland input method, so there's no live preview. "
+            "System Settings → Keyboard → Virtual Keyboard → IBus Wayland, then log out and in.")
+
+
+# What CTranslate2 4.x's CUDA build loads at the first encode.
+_CUDA_LIBS = ("libcublas.so.12", "libcudnn.so.9")
+
+
+def _gpu_libs_warning() -> str:
+    """Why Whisper can't use the GPU it sees, when that's a missing CUDA
+    library (a torch for another CUDA major, pulled in by the train extra,
+    is the usual way). The daemon then decodes on CPU instead."""
+    try:
+        from voiceio.config import load
+        cfg = load().model
+        from voiceio.core.decoders.catalog import is_whisper
+        if cfg.device not in ("auto", "cuda") or not is_whisper(cfg.name):
+            return ""
+        import ctranslate2
+        if ctranslate2.get_cuda_device_count() == 0:
+            return ""
+    except Exception:  # noqa: BLE001 — no faster-whisper, or no CUDA build
+        return ""
+    import ctypes
+    missing = []
+    for lib in _CUDA_LIBS:
+        try:
+            ctypes.CDLL(lib)
+        except OSError:
+            missing.append(lib)
+    if not missing:
+        return ""
+    return (f"A GPU is present but {', '.join(missing)} can't be loaded, so Whisper "
+            "decodes on CPU. Set model.device = \"cpu\" in config.toml to stay on CPU, "
+            "or install the CUDA 12 runtime (cuBLAS 12 and cuDNN 9) to use the GPU.")
+
+
 def _check_features() -> FeatureGroups:
     """Check which features are active based on config and files."""
     groups = FeatureGroups()

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/platform.py

@@ -15,6 +15,22 @@ log = logging.getLogger(__name__)
 DESKTOP_EXTRA_INSTALL = "pipx install --force 'python-voiceio[desktop]'"
 
 
+class LazySounddevice:
+    """`sounddevice`, imported on first use. Its import initializes PortAudio,
+    which raises when libpulse is installed but no sound server runs (CI,
+    headless boxes): imported eagerly, that broke `import voiceio.app` and
+    every command behind it."""
+
+    def __getattr__(self, name: str):
+        if name.startswith("_"):  # introspection (mock, inspect) must not import it
+            raise AttributeError(name)
+        try:
+            import sounddevice
+        except ImportError as e:
+            raise ImportError(desktop_dep_hint("sounddevice")) from e
+        return getattr(sounddevice, name)
+
+
 def desktop_dep_hint(package: str) -> str:
     """One-line message for a desktop dependency that failed to import."""
     return (

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/postprocess.py

@@ -19,8 +19,9 @@ _NO_CASE_LANGUAGES = frozenset({"zh", "ja", "ko", "ar", "he", "th", "hi", "bn",
 #     own line doesn't swallow the paragraph/list break around it.
 # Word repetitions ("had had" is valid English) and filler "like"/"you know"
 # need judgment and are left to the LLM layer. Order matters: multi-token
-# "uh-huh" before "[Uu]h+" so it isn't clipped to a stray "-huh".
-_FILLER = r"\b(?:[Uu]h[-\s]?huh|[Mm]hm|[Uu]h+m*|[Uu]m+|[Hh]m{2,})\b"
+# "uh-huh" and "mm-hmm" (Parakeet invents "Mm-hmm." on a clip's quiet tail)
+# before the bare forms. Never a piece of a hyphenated word: that left "Mm-.".
+_FILLER = r"(?<!-)\b(?:[Uu]h[-\s]?huh|[Mm]m-?[Hh]m+|[Mm]hm|[Uu]h+m*|[Uu]m+|[Hh]m{2,})\b(?!-)"
 # A run of fillers and the comma/space around it; the lowercase letter after it, when
 # it opened a sentence, gets capitalized ("those. Um so I" -> "those. So I"),
 # or the sentence end before it would read as a false stop and become a comma.

{python_voiceio-1.15.1 → python_voiceio-1.16.0}/voiceio/recorder.py

@@ -8,19 +8,15 @@ from typing import TYPE_CHECKING, Callable
 
 import numpy as np
 
-try:
-    import sounddevice as sd
-except ImportError as _e:
-    from voiceio.platform import desktop_dep_hint
-    raise ImportError(desktop_dep_hint("sounddevice")) from _e
-
 from voiceio.audio_source import AudioBuffer
+from voiceio.platform import LazySounddevice
 
 if TYPE_CHECKING:
     from voiceio.config import AudioConfig
     from voiceio.vad import VadBackend
 
 log = logging.getLogger(__name__)
+sd = LazySounddevice()
 
 # Peak amplitude below this is treated as no signal at all (a muted mic emits
 # exact zeros; a working mic's "silence" still noises above ~1e-4). Matches the