python-voiceio 1.8.2__py3-none-any.whl → 1.8.3__py3-none-any.whl

{python_voiceio-1.8.2.dist-info → python_voiceio-1.8.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 1.8.2
+Version: 1.8.3
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-1.8.2.dist-info → python_voiceio-1.8.3.dist-info}/RECORD

@@ -1,5 +1,5 @@
-python_voiceio-1.8.2.dist-info/licenses/LICENSE,sha256=Gz61o8jFTAvZUZyB3nWDB3DQQVuipjfPkVu9W8hBHM0,1072
-voiceio/__init__.py,sha256=d6593s-XBNvVxri9lr2qLUDZQ3Zk3-VXHEwdb4pj8qA,22
+python_voiceio-1.8.3.dist-info/licenses/LICENSE,sha256=Gz61o8jFTAvZUZyB3nWDB3DQQVuipjfPkVu9W8hBHM0,1072
+voiceio/__init__.py,sha256=8wMM-82fXotDjTthMLVtQUtYqY8Z1cIWh3uJI0OjV3s,22
 voiceio/__main__.py,sha256=xT5QCGGreYMisHO7Lh_Y-xAQ2TOzG2D7npKknGNcWSY,53
 voiceio/app.py,sha256=kOvdXNDovKInC1B5jCNx1iyjasFzLTQy_ZEizEF3YkI,42268
 voiceio/audio_source.py,sha256=rR6NRtNzmZ8uqmNPxareBWTJmuBaOxIGI5QINR_l-qc,9102
@@ -23,7 +23,7 @@ voiceio/numbers.py,sha256=CXJtlMhmV6_9Z1FtKAF4-RXycw8voV-BYHPBpdw1I6Y,13571
 voiceio/pidlock.py,sha256=OZLNskUyv9_F3WgyHZWw8WmKt0W7ioxH6xgHitOwCww,1216
 voiceio/platform.py,sha256=uG_5KeYey_29GCn4nSOQJ1UpoCrBC_m2DR7IeoRnuL4,19142
 voiceio/postcorrect.py,sha256=JN5l9xnyGfYq2_owSUwkJ513QiQQKxCrT3hjU5aJCpg,17322
-voiceio/postprocess.py,sha256=rzKmf_ZCskhGnqLEbHaiQX-Sz9yutdJwcwposz7GJ4Q,4661
+voiceio/postprocess.py,sha256=9vpth8u2xOIKU6SimY-hsEDrs4X9YJSI67ZO_f2PpxM,6048
 voiceio/power.py,sha256=mKiG5ioHbzgtW2S-OKGFOPzZFbG0HwUUyMn8kE5JRoQ,681
 voiceio/privacy.py,sha256=FqnA9jx9RCqWW-i6Dh7JcT7vzz97kbUGyQEvsHLvpqc,573
 voiceio/prompt.py,sha256=9iza8KoQxvO_U9NPMJ8Ep7j-zkIs9nW8N5sacCRpsEk,2904
@@ -39,13 +39,13 @@ voiceio/vad.py,sha256=VPeefjJ0Ml5hOuRE2V6BWoBXmzawxLB7if5UP3vBKiM,6030
 voiceio/vocab_stats.py,sha256=md_85Bs0UsM-tmSZDEyQFe0C7z4UnsZ5QSV5vuZ8A5w,5408
 voiceio/vocabulary.py,sha256=i2RRYGmT81tJcYhnZ4sVuVbmZjClqFZVlNUUFa35tuY,11351
 voiceio/watchdog.py,sha256=XnrpM3ULeUoJ-sMt_aNNzJSLqOuK4ocJAMHQ00knYuw,7838
-voiceio/wordfreq.py,sha256=PTAoaEUX1D2Y12jjd8A9cFTqsZkDDZt6eesBBL_4toA,3151
+voiceio/wordfreq.py,sha256=AxmVgpQRfm2YAYQpHCRRrBUTpa12y_ngiFNScrGr4GM,3543
 voiceio/cli/__init__.py,sha256=EM2Qa1COWkqSQ9hFuPKdf8_OD5JoGsY3QCpQA-waEJQ,8353
 voiceio/cli/configcmd.py,sha256=YF7t-LkwEYEGH4qcO86xehWVX0LvltqxjDFpMF5ljjA,2858
 voiceio/cli/corrections.py,sha256=OBjgiOhtWlTQpQq0HLAvk5-EfKJ7NNt-hnNBHm7obEc,4247
 voiceio/cli/doctor.py,sha256=P8bOn4xedMistlKHhNeeGxRsmogV6jBVKc-CUbo0LXs,7769
 voiceio/cli/history.py,sha256=7FvEEgdw8jf5XXvC7-4S7iIpcjBbn7dx4czIijg61C0,2597
-voiceio/cli/learn.py,sha256=9i1mW4XUbSN74_afZjt1gbL77flx8WzctLHW0oKHArs,32300
+voiceio/cli/learn.py,sha256=ajVP7RRFBLXa0Sf-pbYrccPpuB58oHjd4wR9JkcOwtw,32914
 voiceio/cli/models.py,sha256=fye_mHbsXLY11uRC1BCDlucCHPny9ftoehAeUJPMJv0,4263
 voiceio/cli/servicecmd.py,sha256=2St2WfLkt06EpC51I5i9Ykv_pvLKzu_dz01M-29O3-4,4830
 voiceio/cli/uninstall.py,sha256=Vqnq6-4hjhDrxqkU_1wtXkY5BU_Tqa-TjZA0Qpf12e4,8921
@@ -85,7 +85,7 @@ voiceio/learn/metrics.py,sha256=b2xWUzPpNpXHn6zgS1upPbEcpW3ExH9SmSLIUzPhStQ,5174
 voiceio/learn/quiet.py,sha256=5Ov3mvE1G3CjhwFQOuJvGeju242zFASi29tketnUyew,1523
 voiceio/learn/review.py,sha256=X5qs5yEf20Ao3gMhsbW3Eur62hSqy9AYlv08Pcrllxo,5572
 voiceio/learn/store.py,sha256=MauWhGaI2drMbiChIZOblpPS_uVAw-1sNWGyGACHYw8,12816
-voiceio/learn/teacher.py,sha256=uF4hQkrdjv012TTBRSwrhQqqAeCFNoidABcrgzSVUoI,3696
+voiceio/learn/teacher.py,sha256=FhCv08Js3chs4Vbh0VN3ZI9p05HFxeeeFJF8jTjjzPQ,3751
 voiceio/learn/train.py,sha256=FtZMa4ddsPMRUxkyqC-Qf9OveZmgRMHiGXx5WHj_jT8,18214
 voiceio/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 voiceio/models/silero_vad.onnx,sha256=GhU6IvRQnikqlOZ9b5uF6N6yW0mIaCt-F0xlJ52HiOM,2327524
@@ -132,8 +132,8 @@ voiceio/typers/pynput_type.py,sha256=f3w131YWJfNGd4IAcWtuKd2Gcwu3467QwGW3OBUEXNI
 voiceio/typers/wtype.py,sha256=11vOAKHjMat-fITd4sUoUWwgniyVp65EbH3i4mL6Vhg,1762
 voiceio/typers/xdotool.py,sha256=OuT_zW_z07jSAqu9Cmn_gL1mJfQsiX62ecqFRvpAarQ,1332
 voiceio/typers/ydotool.py,sha256=oY5EXB-nI7PvxTL_5Tb3JjfvyhmLNYNnUFvkb1XUn_U,5978
-python_voiceio-1.8.2.dist-info/METADATA,sha256=2SK3PktKq9k4l5bG9ki71XPCxzDES_Nt4SIzJZtjoFk,7519
-python_voiceio-1.8.2.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
-python_voiceio-1.8.2.dist-info/entry_points.txt,sha256=iDIDmhvQ3xeN_CWsFne1t9STZ08I25I_wp0xmde_vMs,94
-python_voiceio-1.8.2.dist-info/top_level.txt,sha256=piwtn309lD6uexQyXdZ-efAVBJF9y6Wfr48Z-8zkNhg,8
-python_voiceio-1.8.2.dist-info/RECORD,,
+python_voiceio-1.8.3.dist-info/METADATA,sha256=Eeq7y0M9-FMTabCANoqNTbMuqXGPCNBUZ8VN_3q1lN0,7519
+python_voiceio-1.8.3.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
+python_voiceio-1.8.3.dist-info/entry_points.txt,sha256=iDIDmhvQ3xeN_CWsFne1t9STZ08I25I_wp0xmde_vMs,94
+python_voiceio-1.8.3.dist-info/top_level.txt,sha256=piwtn309lD6uexQyXdZ-efAVBJF9y6Wfr48Z-8zkNhg,8
+python_voiceio-1.8.3.dist-info/RECORD,,

voiceio/__init__.py

@@ -1 +1 @@
-__version__ = "1.8.2"
+__version__ = "1.8.3"

voiceio/cli/learn.py

@@ -417,6 +417,16 @@ def cmd_train(args: argparse.Namespace) -> None:
               f"  (compare with the current model, then `voiceio learn promote {run.name}`)")
 
 
+def _train_language() -> str:
+    """The language a fine-tune is trained under: the one dictation decodes
+    with. "auto" has no single answer, so it has to be set first."""
+    lang = _cfg().model.language or "en"
+    if lang == "auto":
+        raise SystemExit("Fine-tuning needs one language: set it with "
+                         "`voiceio config set model.language es` (your dictation language).")
+    return lang
+
+
 def _train(ds: Path, *, base: str = "small", epochs: int = 4, synthetic_share: float = 0.3,
            full: bool = False, remote: bool = False) -> Path | None:
     """Export `ds` and fine-tune on it; returns the run dir (None for --remote)."""
@@ -433,8 +443,11 @@ def _train(ds: Path, *, base: str = "small", epochs: int = 4, synthetic_share: f
         return None
     train.require_train_deps()
     hotwords = (folder / train.HOTWORDS_FILE).read_text(encoding="utf-8")
+    # The language token the model is trained under must be the one it decodes
+    # with: training a Spanish voice under "en" taught it the wrong prompt.
     hf = train.train(folder, run, train.TrainConfig(
-        base=base, epochs=epochs, decoder_only=False if full else None, hotwords=hotwords))
+        base=base, epochs=epochs, decoder_only=False if full else None, hotwords=hotwords,
+        language=_train_language()))
     train.convert(hf, run / "ct2")
     return run
 

voiceio/learn/teacher.py

@@ -45,7 +45,7 @@ class WhisperTeacher:
         # A background batch job (see cli.learn._yield_cpu): use every core idle.
         self._model = load_model(model, device=device, compute_type=compute_type,
                                  cpu_threads=max(4, (os.cpu_count() or 8) // 2))
-        self._language = language
+        self._language = None if language in ("", "auto") else language   # None: detect
         terms = [t for t in vocabulary if t]
         # Prompt, not hotwords: hotwords are capped at 223 tokens; a prompt
         # of the most important terms still biases spelling. The tokenizer

voiceio/postprocess.py

@@ -23,6 +23,21 @@ _NO_CASE_LANGUAGES = frozenset({"zh", "ja", "ko", "ar", "he", "th", "hi", "bn",
 _FILLER_RE = re.compile(
     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 word (2+ letters, not part of a dotted name or an ellipsis) ending in "."
+# and followed by a lowercase word: a segment boundary, not a sentence end.
+_SEGMENT_STOP_RE = re.compile(r"(?<![.\w])([A-Za-zÀ-ɏ]{2,})\.(\s+)(?=[a-zà-ÿ])")
+_ABBREVIATIONS = frozenset({
+    "etc", "vs", "approx", "cf", "al", "mr", "mrs", "ms", "dr", "st", "jr", "sr",
+    "inc", "ltd", "co", "fig", "eq", "vol", "no", "pp", "ca", "min", "max", "sec",
+})
+
+
+def _not_an_ending(word: str) -> bool:
+    """The word before a "." is an abbreviation or a dotted name, not a
+    sentence's last word."""
+    return "." in word or word.lower() in _ABBREVIATIONS
+
+
 # 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
@@ -95,14 +110,28 @@ def cleanup(text: str, language: str = "en") -> str:
     text = re.sub(r"\s+([.,;:?!])", r"\1", text)
 
     if language not in _NO_CASE_LANGUAGES:
+        # Whisper closes every segment with "." even mid-sentence, and the next
+        # segment then carries on in lowercase ("a bit bigger. and remove the
+        # latch"): on real dictation that was 1 in 4 notes. A lowercase
+        # continuation means the sentence did not end, so it gets a comma —
+        # before the capitalization below would make the false stop permanent.
+        text = _SEGMENT_STOP_RE.sub(
+            lambda m: m.group(0) if m.group(1).lower() in _ABBREVIATIONS
+            else m.group(1) + "," + m.group(2),
+            text,
+        )
+
+
         # Capitalize first character
         text = text[0].upper() + text[1:]
 
         # Capitalize after sentence-ending punctuation (an ellipsis is a
-        # pause, not an ending)
+        # pause, not an ending, and so is the dot of "etc.", "e.g." or a name
+        # like "hugo0.com")
         text = re.sub(
-            r"(?<!\.\.)([.?!]\s+)([a-z\u00E0-\u00FF])",
-            lambda m: m.group(1) + m.group(2).upper(),
+            r"(\S*)(?<!\.\.)([.?!]\s+)([a-z\u00E0-\u00FF])",
+            lambda m: m.group(0) if m.group(2)[0] == "." and _not_an_ending(m.group(1))
+            else m.group(1) + m.group(2) + m.group(3).upper(),
             text,
         )
 

voiceio/wordfreq.py

@@ -41,14 +41,26 @@ _CONTRACTIONS = frozenset({
 })
 
 
+def _zipf(word: str, language: str) -> float:
+    """Zipf frequency, 0.0 for a language wordfreq has no list for. "auto"
+    (detect per dictation) has no single list either: score it as English,
+    which is where the jargon in a vocabulary list usually comes from."""
+    if language in ("", "auto"):
+        language = "en"
+    try:
+        return zipf_frequency(word.lower(), language)
+    except LookupError:
+        return 0.0
+
+
 def is_common(word: str, language: str = "en", threshold: float = _ZIPF_COMMON_THRESHOLD) -> bool:
     """Check if a word is common (high frequency). Used for suggesting corrections."""
-    return zipf_frequency(word.lower(), language) >= threshold
+    return _zipf(word, language) >= threshold
 
 
 def rank(word: str, language: str = "en") -> int | None:
     """Get approximate frequency rank. None if not found."""
-    z = zipf_frequency(word.lower(), language)
+    z = _zipf(word, language)
     if z == 0.0:
         return None
     # Convert zipf to approximate rank (higher zipf = lower rank number)
@@ -61,7 +73,7 @@ def is_known(word: str, language: str = "en") -> bool:
     wl = word.lower()
     if wl in _CONTRACTIONS:
         return True
-    return zipf_frequency(wl, language) >= _ZIPF_KNOWN_THRESHOLD
+    return _zipf(wl, language) >= _ZIPF_KNOWN_THRESHOLD
 
 
 def extract_words(text: str) -> list[str]: