python-voiceio 0.9.6__tar.gz → 0.9.8__tar.gz

{python_voiceio-0.9.6/python_voiceio.egg-info → python_voiceio-0.9.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 0.9.6
+Version: 0.9.8
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "python-voiceio"
-version = "0.9.6"
+version = "0.9.8"
 description = "Voice dictation for Linux. Speak → text, locally, instantly."
 readme = "README.md"
 license = "MIT"

{python_voiceio-0.9.6 → python_voiceio-0.9.8/python_voiceio.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 0.9.6
+Version: 0.9.8
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/tests/test_platform.py

@@ -72,3 +72,34 @@ def test_platform_frozen():
     import pytest
     with pytest.raises(AttributeError):
         p.os = "darwin"
+
+
+class TestOpenInTerminal:
+    """The tray menu launches print-and-exit commands (history/doctor/logs);
+    hold=True must keep the window open instead of flashing it shut."""
+
+    def _capture(self, monkeypatch, *, hold):
+        import voiceio.platform as plat
+        monkeypatch.setattr(plat, "_detect_os", lambda: "linux")
+        monkeypatch.setenv("TERMINAL", "myterm")
+        monkeypatch.setattr(
+            plat.shutil, "which",
+            lambda b: "/usr/bin/myterm" if b == "myterm" else None,
+        )
+        calls = []
+        with patch("subprocess.Popen", side_effect=lambda args, *a, **k: calls.append(args)):
+            ok = plat.open_in_terminal(["voiceio", "history"], hold=hold)
+        assert ok is True
+        return calls[0]
+
+    def test_hold_wraps_in_shell_that_pauses(self, monkeypatch):
+        args = self._capture(monkeypatch, hold=True)
+        assert args[:2] == ["myterm", "-e"]
+        assert "sh" in args and "-c" in args
+        script = args[-1]
+        assert "voiceio history" in script
+        assert "read" in script  # pauses so the window survives
+
+    def test_no_hold_runs_command_directly(self, monkeypatch):
+        args = self._capture(monkeypatch, hold=False)
+        assert args == ["myterm", "-e", "voiceio", "history"]

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/tests/test_postcorrect.py

@@ -121,6 +121,25 @@ def test_disfluency_mode_rejects_rewording():
         assert pc.correct(original) == original
 
 
+def test_disfluency_mode_rejects_dropped_negation():
+    """Dropping 'not' is one deletion under every fraction cap but inverts
+    meaning — the negation guard must reject it."""
+    pc = PostCorrector(_cfg(remove_disfluencies=True))
+    original = "I do not want the feature to ship today at all"
+    inverted = "I do want the feature to ship today at all"
+    with patch("voiceio.llm_api.chat", return_value=inverted):
+        assert pc.correct(original) == original
+
+
+def test_disfluency_mode_rejects_replaced_negation():
+    """Replacing a negation with something else also inverts meaning."""
+    pc = PostCorrector(_cfg(remove_disfluencies=True))
+    original = "we should never deploy on a friday afternoon"
+    changed = "we should always deploy on a friday afternoon"
+    with patch("voiceio.llm_api.chat", return_value=changed):
+        assert pc.correct(original) == original
+
+
 def test_fix_mode_still_rejects_big_deletion():
     """With disfluency mode OFF, the original conservative guards stand: a big
     deletion is a length change and must be rejected."""

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/tests/test_postprocess.py

@@ -24,15 +24,48 @@ class TestStripDisfluencies:
         assert strip_disfluencies("I had had enough") == "I had had enough"
 
     def test_dedups_duplicate_sentence(self):
-        # The Whisper re-decode artifact: an exact sentence repeated.
-        out = strip_disfluencies("Do the research. Do the research.")
-        assert out == "Do the research."
+        # The Whisper re-decode artifact: a whole sentence repeated verbatim.
+        out = strip_disfluencies("Do the deep research now. Do the deep research now.")
+        assert out == "Do the deep research now."
 
     def test_keeps_meaningful_words(self):
         # 'like' as a real verb/preposition and content must survive.
         text = "I like the design and it works like a charm"
         assert strip_disfluencies(text) == text
 
+    def test_never_eats_real_words_or_units(self):
+        # Default-on runs on everyone's speech: filler patterns must not collide
+        # with real words, units, or abbreviations. Case-sensitive matching is
+        # what protects the all-caps abbreviations (ER, UM, HM).
+        for text in [
+            "to err is human",
+            "we should err on caution",
+            "the bolt is 5 mm wide",
+            "set it to 10 mm please",
+            "ah yes I remember now",
+            "I like the ohm rating",
+            "Take him to the ER right now",       # ER = emergency room
+            "The UM campus in Michigan",           # UM = University of Michigan
+            "The Er atom is a lanthanide",         # Er = erbium
+            "we measured 3 hm across",             # hm = hectometre (bare, 1 m)
+        ]:
+            assert strip_disfluencies(text) == text, text
+
+    def test_preserves_newlines_and_structure(self):
+        # A filler on its own line must not swallow the paragraph break.
+        assert strip_disfluencies("First para.\n\nSecond para.") == \
+            "First para.\n\nSecond para."
+        assert "\n\n" in strip_disfluencies("First para.\n\num\n\nSecond para.")
+
+    def test_uh_huh_removed_whole(self):
+        # Regression: ordering bug once stranded "-huh".
+        assert strip_disfluencies("uh-huh right") == "right"
+        assert strip_disfluencies("uh huh yes") == "yes"
+
+    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."
+
     def test_empty(self):
         assert strip_disfluencies("") == ""
 

python_voiceio-0.9.8/voiceio/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.9.8"

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/voiceio/config.py

@@ -93,7 +93,8 @@ class OutputConfig:
     # and duplicate re-decode sentences on every pass; when [postcorrect] is
     # also enabled, its LLM pass additionally removes false starts and filler
     # "like", guarded so it can only delete — never add, rephrase, or reorder.
-    remove_disfluencies: bool = False
+    # On by default: nobody wants "um" in their notes. Set false for verbatim.
+    remove_disfluencies: bool = True
     voice_input_prefix: str = ""           # e.g. "[voice]" — empty disables
     # Incremental finalization: once the un-finalized audio tail grows past
     # this many seconds, it is cut at the nearest interior speech pause,

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/voiceio/platform.py

@@ -311,8 +311,14 @@ def system_deps_install_cmd(keys: list[str] | None = None) -> str:
     return f"{prefix} {' '.join(pkgs)}"
 
 
-def open_in_terminal(cmd: list[str]) -> bool:
-    """Launch a command in the user's terminal emulator. Returns success."""
+def open_in_terminal(cmd: list[str], *, hold: bool = False) -> bool:
+    """Launch a command in the user's terminal emulator. Returns success.
+
+    ``hold=True`` keeps the window open after the command exits — required for
+    print-and-exit commands (history, doctor, logs) that would otherwise flash
+    a terminal that closes the instant they finish.
+    """
+    import shlex
     import subprocess
 
     plat_os = _detect_os()
@@ -334,11 +340,20 @@ def open_in_terminal(cmd: list[str]) -> bool:
             log.warning("Failed to open cmd.exe")
             return False
 
-    # Linux: try common terminal emulators
+    # Linux: try common terminal emulators. When holding, wrap the command in a
+    # shell that pauses at the end so the window survives a print-and-exit.
+    launch = cmd
+    if hold:
+        inner = " ".join(shlex.quote(c) for c in cmd)
+        launch = [
+            "sh", "-c",
+            f'{inner}; printf "\\n[voiceio] Press Enter to close…"; read _',
+        ]
+
     term = os.environ.get("TERMINAL")
     if term and shutil.which(term):
         try:
-            subprocess.Popen([term, "-e", *cmd])
+            subprocess.Popen([term, "-e", *launch])
             return True
         except OSError:
             pass
@@ -346,7 +361,7 @@ def open_in_terminal(cmd: list[str]) -> bool:
     # x-terminal-emulator (Debian/Ubuntu alternative)
     if shutil.which("x-terminal-emulator"):
         try:
-            subprocess.Popen(["x-terminal-emulator", "-e", *cmd])
+            subprocess.Popen(["x-terminal-emulator", "-e", *launch])
             return True
         except OSError:
             pass
@@ -365,7 +380,7 @@ def open_in_terminal(cmd: list[str]) -> bool:
     for prefix, binary in _TERMINALS:
         if shutil.which(binary):
             try:
-                subprocess.Popen([*prefix, *cmd])
+                subprocess.Popen([*prefix, *launch])
                 return True
             except OSError:
                 continue

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/voiceio/postcorrect.py

@@ -34,6 +34,19 @@ _MAX_INSERTED_WORDS = 0     # adding any word = altering meaning → reject
 _MAX_REPLACE_FRAC = 0.15    # ASR word-fixes only, never wholesale rewording
 _MAX_DELETE_FRAC = 0.4      # backstop against deleting real content
 
+# Words whose deletion/replacement flips meaning — the fraction caps can't catch
+# a single dropped "not". If the edit touches any of these on the original side,
+# reject outright. (Contractions ending in "n't" are handled separately.)
+_MEANING_CRITICAL = frozenset({
+    "not", "no", "never", "none", "nor", "neither", "without", "cannot",
+    "nothing", "nobody", "nowhere", "n't",
+})
+
+
+def _is_meaning_critical(word: str) -> bool:
+    w = word.lower().strip(".,;:!?\"'()")
+    return w in _MEANING_CRITICAL or w.endswith("n't")
+
 _SYSTEM_PROMPT = (
     "You fix automatic speech recognition errors in dictated text. "
     "The user dictates about software engineering and their projects. "
@@ -46,18 +59,25 @@ _SYSTEM_PROMPT = (
 # Disfluency mode: also strip spoken filler, delete-only. The strict rules
 # mirror the guards — the model is told exactly what the diff check enforces.
 _SYSTEM_PROMPT_CLEAN = (
-    "You convert dictated speech into clean written text. The user dictates "
-    "about software engineering and their projects. Do exactly two things:\n"
+    "You convert dictated speech into clean, concise written text. The user "
+    "dictates about software engineering and their projects, and wants it to "
+    "read like something they wrote, not spoke. Do exactly two things:\n"
     "1. Fix words the recognizer misheard (wrong proper nouns, homophones, "
     "garbled technical terms).\n"
-    "2. Remove speech disfluencies: filler sounds (um, uh, er); filler uses of "
-    "'like', 'you know', 'I mean'; false starts and self-corrections (keep the "
-    "corrected version); and stray word repetitions.\n"
-    "STRICT RULES: Only DELETE disfluencies and FIX misheard words. NEVER add "
-    "words. NEVER rephrase, reword, reorder, or summarize. NEVER drop real "
-    "content, meaningful hedges, or negations. If unsure whether something is a "
-    "disfluency, KEEP it. Preserve the speaker's own wording and punctuation. "
-    "Return only the cleaned text, nothing else."
+    "2. Cut spoken disfluency so it reads tight and clear:\n"
+    "   - filler sounds (um, uh, er);\n"
+    "   - filler discourse markers with no content: leading 'so', 'yeah', "
+    "'okay', 'well', 'now', 'right', and 'you know', 'I mean', 'like', 'sort "
+    "of'/'kind of' when used as filler, 'or something'/'or whatever', a "
+    "trailing 'right?' tag;\n"
+    "   - false starts and self-corrections (keep only the corrected version);\n"
+    "   - stray word and phrase repetitions.\n"
+    "Lean toward the briefer version — remove hesitation and padding freely.\n"
+    "HARD LIMITS (these protect meaning): only DELETE and FIX — NEVER add "
+    "words, rephrase, reword, reorder, or summarize. NEVER drop or alter actual "
+    "content, a negation ('not', 'never', \"n't\"), or a hedge that changes "
+    "certainty ('I think', 'maybe', 'probably', 'might'). Keep the speaker's own "
+    "wording. Return only the cleaned text, nothing else."
 )
 
 _MAX_RECENT = 3
@@ -315,8 +335,12 @@ class PostCorrector:
                 inserted += j2 - j1
             elif tag == "replace":
                 replaced += max(i2 - i1, j2 - j1)
+                if any(_is_meaning_critical(w) for w in aw[i1:i2]):
+                    return False, "negation"   # e.g. "not" → something else
             elif tag == "delete":
                 deleted += i2 - i1
+                if any(_is_meaning_critical(w) for w in aw[i1:i2]):
+                    return False, "negation"   # dropping "not" inverts meaning
         n = len(aw)
         if inserted > _MAX_INSERTED_WORDS:
             return False, "inserted"          # added content — meaning changed

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/voiceio/postprocess.py

@@ -14,13 +14,25 @@ if TYPE_CHECKING:
 _NO_CASE_LANGUAGES = frozenset({"zh", "ja", "ko", "ar", "he", "th", "hi", "bn", "ka", "my"})
 
 # Filler SOUNDS only — tokens with no lexical meaning, so deleting them can
-# never change meaning. Word repetitions ("had had" is valid English) and
-# filler uses of "like"/"you know" need judgment and are left to the LLM layer.
-# Consumes an adjacent comma on either side so "be, uh, found" → "be found".
+# never change meaning. Deliberately conservative because this runs on every
+# user's speech:
+#   * CASE-SENSITIVE (no IGNORECASE): all-caps abbreviations that look like
+#     fillers are real words and must survive — "ER" (emergency room), "UM"
+#     (University of Michigan), "HM". We match lowercase and Title-case forms
+#     ("um", "Um") — Whisper's filler spellings — but never all-caps.
+#   * Excluded entirely: "er"/"erm" ("ER"/"Er"=erbium), "mm" (millimetres),
+#     "ah" (interjection), bare "hm" (hectometre — require "hmm", 2+ m's).
+#   * Surrounding whitespace is horizontal-only ([^\S\n]) so a filler on its
+#     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_RE = re.compile(
-    r"\s*,?\s*\b(?:u+m+|u+h+m*|e+r+m*|erm+|a+h+|h+m+|mm+|mhm|uh[-\s]?huh)\b\s*,?\s*",
-    re.IGNORECASE,
+    r"[^\S\n]*,?[^\S\n]*\b(?:[Uu]h[-\s]?huh|[Mm]hm|[Uu]h+m*|[Uu]m+|[Hh]m{2,})\b[^\S\n]*,?[^\S\n]*",
 )
+# A re-decode artifact is a whole duplicated sentence; require this many words
+# so emphatic short repeats ("No. No.", "Stop. Stop.") are preserved.
+_MIN_DEDUP_WORDS = 4
 
 
 def strip_disfluencies(text: str) -> str:
@@ -36,19 +48,33 @@ def strip_disfluencies(text: str) -> str:
         return text
     text = _FILLER_RE.sub(" ", text)
     text = _dedup_adjacent_sentences(text)
-    text = re.sub(r"\s{2,}", " ", text)
+    # Repair the debris the deletions leave, without crossing newlines (so
+    # paragraph/list structure survives even when punctuation_cleanup is off).
+    text = re.sub(r"[^\S\n]+([,.;:?!])", r"\1", text)  # space before punctuation
+    text = re.sub(r"[^\S\n]{2,}", " ", text)           # collapse runs of spaces
     return text.strip()
 
 
 def _dedup_adjacent_sentences(text: str) -> str:
-    """Drop a sentence identical to the one immediately before it."""
-    parts = re.split(r"(?<=[.?!])\s+", text)
+    """Drop a sentence identical to the one immediately before it.
+
+    Separators are captured and preserved on rejoin so paragraph/list breaks
+    survive; only a full (>= _MIN_DEDUP_WORDS) verbatim repeat is removed.
+    """
+    tokens = re.split(r"((?<=[.?!])\s+)", text)  # [sent, sep, sent, sep, …]
     out: list[str] = []
-    for p in parts:
-        if out and p.strip().lower() == out[-1].strip().lower():
-            continue
-        out.append(p)
-    return " ".join(out)
+    last_kept: str | None = None
+    for i in range(0, len(tokens), 2):
+        sentence = tokens[i]
+        sep = tokens[i + 1] if i + 1 < len(tokens) else ""
+        norm = sentence.strip().lower()
+        if (last_kept is not None and norm == last_kept
+                and len(sentence.split()) >= _MIN_DEDUP_WORDS):
+            continue  # drop the duplicate sentence and its separator
+        out.append(sentence)
+        out.append(sep)
+        last_kept = norm
+    return "".join(out)
 
 
 def cleanup(text: str, language: str = "en") -> str:

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/voiceio/tray/__init__.py

@@ -146,7 +146,7 @@ def _read_stdout(proc: subprocess.Popen, toggle_cb: Callable[[], None]) -> None:
                 action = cmd[5:]
                 cli_cmd = _MENU_COMMANDS.get(action)
                 if cli_cmd:
-                    if not open_in_terminal(cli_cmd):
+                    if not open_in_terminal(cli_cmd, hold=True):
                         log.warning("Failed to open terminal for: %s", action)
     except (OSError, ValueError):
         pass

{python_voiceio-0.9.6 → python_voiceio-0.9.8}/voiceio/tray/_pystray.py

@@ -45,7 +45,7 @@ def start(
             from voiceio.platform import open_in_terminal
             cli_cmd = _MENU_COMMANDS.get(action)
             if cli_cmd:
-                open_in_terminal(cli_cmd)
+                open_in_terminal(cli_cmd, hold=True)
         return _handler
 
     _icon = pystray.Icon(

python_voiceio-0.9.6/voiceio/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.9.6"