legendaria 0.1.0.dev2__tar.gz → 0.1.0.dev3__tar.gz

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: legendaria
-Version: 0.1.0.dev2
+Version: 0.1.0.dev3
 Summary: Legendária transcription engine: local Whisper transcription with word timings (CLI: flosub).
 Project-URL: Homepage, https://github.com/flou-ainan/Legendaria
 Project-URL: Issues, https://github.com/flou-ainan/Legendaria/issues

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/legendaria_engine/__init__.py

@@ -1,3 +1,3 @@
 """Legendária transcription engine."""
 
-__version__ = "0.1.0.dev2"
+__version__ = "0.1.0.dev3"

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/legendaria_engine/config.py

@@ -16,6 +16,8 @@ STYLE_NUMBERS = (
     "max_duration",
     "min_duration",
     "pause_break",
+    "join_short",
+    "fill_gaps",
 )
 
 

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/legendaria_engine/cues.py

@@ -8,6 +8,7 @@ from __future__ import annotations
 
 import json
 from dataclasses import asdict, dataclass, field
+from itertools import pairwise
 from pathlib import Path
 from typing import Any
 
@@ -56,12 +57,30 @@ def _ends_with(word: Word, chars: str) -> bool:
     return bool(text) and bool(chars) and text[-1] in chars
 
 
+def _is_attached(word: Word) -> bool:
+    """A piece Whisper glued to the previous word (no leading space): "-vindas", "'s"."""
+    return bool(word.word) and not word.word[0].isspace()
+
+
+def _units(words: list[Word]) -> int:
+    """Readable words: attached pieces do not count ("bem" + "-vindas" is one)."""
+    return 1 + sum(1 for w in words[1:] if not _is_attached(w)) if words else 0
+
+
+def _letters(words: list[Word]) -> int:
+    return sum(ch.isalnum() for ch in _join(words))
+
+
 def make_cues(transcript: Transcript, style: dict[str, Any], style_name: str = "custom") -> CueList:
     """Chunk the transcript's words into cues that respect the style's limits.
 
-    Style keys (all optional): max_words, max_chars, max_duration, min_duration,
-    pause_break (seconds of silence that force a break), break_after (punctuation
-    characters that end a cue), break_on_segment (start a new cue at each segment).
+    Style keys (all optional): max_words (readable words; attached pieces like "-vindas"
+    never split from their word), max_chars, max_duration, min_duration, pause_break
+    (seconds of silence that force a break), break_after (punctuation characters that
+    end a cue), break_on_segment (start a new cue at each segment), join_short (a single
+    word with at most this many letters, like "a", "de", "no", may take the next word
+    even beyond max_words), fill_gaps (extend a cue to the next one when the silence
+    between them is at most this many seconds, so the screen does not blink).
     """
     max_words = style.get("max_words")
     max_chars = style.get("max_chars")
@@ -69,6 +88,7 @@ def make_cues(transcript: Transcript, style: dict[str, Any], style_name: str = "
     pause_break = style.get("pause_break")
     break_after = style.get("break_after", "")
     break_on_segment = style.get("break_on_segment", True)
+    join_short = style.get("join_short")
 
     cues: list[Cue] = []
     current: list[Word] = []
@@ -87,8 +107,22 @@ def make_cues(transcript: Transcript, style: dict[str, Any], style_name: str = "
         if break_on_segment:
             flush()
         for word in segment.words:
+            if current and _is_attached(word):
+                current.append(word)  # glued piece: never starts a new cue
+                if _ends_with(word, break_after):
+                    flush()
+                continue
             if current:
-                too_many = max_words is not None and len(current) + 1 > max_words
+                allowed = max_words
+                if (
+                    allowed is not None
+                    and join_short
+                    and _units(current) == 1
+                    and _letters(current) <= join_short
+                    and not _ends_with(current[-1], break_after)
+                ):
+                    allowed += 1  # short word ("de", "a") takes the next one
+                too_many = allowed is not None and _units(current) + 1 > allowed
                 too_long = max_chars is not None and len(_join([*current, word])) > max_chars
                 too_slow = max_duration is not None and word.end - current[0].start > max_duration
                 paused = pause_break is not None and word.start - current[-1].end > pause_break
@@ -98,12 +132,22 @@ def make_cues(transcript: Transcript, style: dict[str, Any], style_name: str = "
             if _ends_with(word, break_after):
                 flush()
     flush()
+    _fill_gaps(cues, style.get("fill_gaps"))
     _apply_min_duration(cues, style.get("min_duration"))
     return CueList(
         style=style_name, source=transcript.source, language=transcript.language, cues=cues
     )
 
 
+def _fill_gaps(cues: list[Cue], max_gap: float | None) -> None:
+    """Extend each cue to the next one's start when the silence between them is short."""
+    if not max_gap:
+        return
+    for cue, nxt in pairwise(cues):
+        if 0 < nxt.start - cue.end <= max_gap:
+            cue.end = nxt.start
+
+
 def _apply_min_duration(cues: list[Cue], min_duration: float | None) -> None:
     """Stretch cues shorter than `min_duration`, without overlapping the next cue."""
     if not min_duration:

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/legendaria_engine/defaults/config.json

@@ -13,6 +13,17 @@
   "word_timestamps": true,
   "style": "subtitles",
   "styles": {
+    "pop": {
+      "max_words": 1,
+      "join_short": 3,
+      "max_chars": 24,
+      "max_duration": 1.6,
+      "min_duration": 0.25,
+      "pause_break": 0.6,
+      "fill_gaps": 0.35,
+      "break_after": ".,!?;:…",
+      "break_on_segment": true
+    },
     "subtitles": {
       "max_chars": 84,
       "line_length": 42,

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "legendaria"
-version = "0.1.0.dev2"
+version = "0.1.0.dev3"
 description = "Legendária transcription engine: local Whisper transcription with word timings (CLI: flosub)."
 readme = "README.md"
 license = "MIT"

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/schemas/config.schema.json

@@ -96,7 +96,7 @@
     },
     "styles": {
       "type": "object",
-      "description": "Cue styles: how words are grouped into on-screen text. 'captions' = short CapCut-style phrases (Blender default); 'subtitles' = film-style lines (SRT/VTT default). Add your own.",
+      "description": "Cue styles: how words are grouped into on-screen text. 'pop' = one word at a time (two when the first is very short), for pop animations; 'captions' = short CapCut-style phrases (Blender default); 'subtitles' = film-style lines (SRT/VTT default). Add your own.",
       "additionalProperties": {
         "type": "object",
         "properties": {
@@ -155,6 +155,22 @@
           "break_on_segment": {
             "type": "boolean",
             "description": "Start a new cue at each Whisper segment."
+          },
+          "join_short": {
+            "type": [
+              "integer",
+              "null"
+            ],
+            "minimum": 1,
+            "description": "A single word with at most this many letters (\"a\", \"de\", \"no\") may take the next word, even beyond max_words."
+          },
+          "fill_gaps": {
+            "type": [
+              "number",
+              "null"
+            ],
+            "minimum": 0,
+            "description": "Extend a cue to the next one when the silence between them is at most this many seconds (avoids blinking)."
           }
         },
         "additionalProperties": false

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/tests/test_cues.py

@@ -108,3 +108,53 @@ def test_attached_pieces_stay_glued():
         Word(0.5, 0.7, " Tom's"),
     ]
     assert texts(make_cues(transcript(segment(ws)), {})) == ["Descreve-se o Tom's"]
+
+
+def test_attached_pieces_never_start_a_cue():
+    ws = [
+        Word(0, 0.3, " bem"),
+        Word(0.3, 0.6, "-vindas"),
+        Word(0.7, 1.0, " ao"),
+        Word(1.0, 1.3, " canal"),
+    ]
+    result = make_cues(transcript(segment(ws)), {"max_words": 1})
+    assert texts(result) == ["bem-vindas", "ao", "canal"]
+
+
+def test_short_words_pair_with_the_next():
+    ws = words(
+        ("a", 0, 0.1),
+        ("bolsa", 0.1, 0.5),
+        ("de", 0.5, 0.6),
+        ("crochê", 0.6, 1.0),
+        ("linda", 1.0, 1.4),
+    )
+    result = make_cues(transcript(segment(ws)), {"max_words": 1, "join_short": 3})
+    assert texts(result) == ["a bolsa", "de crochê", "linda"]
+
+
+def test_short_word_ending_a_sentence_does_not_pair():
+    ws = words(("é.", 0, 0.2), ("Agora", 0.3, 0.6))
+    result = make_cues(
+        transcript(segment(ws)), {"max_words": 1, "join_short": 3, "break_after": "."}
+    )
+    assert texts(result) == ["é.", "Agora"]
+
+
+def test_fill_gaps_closes_short_silences_only():
+    ws = words(("um", 0.0, 0.2), ("dois", 0.4, 0.6), ("três", 2.0, 2.2))
+    result = make_cues(transcript(segment(ws)), {"max_words": 1, "fill_gaps": 0.35})
+    assert [(c.start, c.end) for c in result.cues] == [(0.0, 0.4), (0.4, 0.6), (2.0, 2.2)]
+
+
+def test_pop_style_on_defaults():
+    config = load_config()
+    ws = words(
+        *[
+            (w, i * 0.3, i * 0.3 + 0.25)
+            for i, w in enumerate(["e", "hoje", "eu", "trouxe", "a", "bolsa"])
+        ]
+    )
+    result = make_cues(transcript(segment(ws)), config["styles"]["pop"], "pop")
+    assert texts(result) == ["e hoje", "eu trouxe", "a bolsa"]
+    assert all(len(c.text.split()) <= 2 for c in result.cues)

{legendaria-0.1.0.dev2 → legendaria-0.1.0.dev3}/uv.lock

@@ -472,7 +472,7 @@ wheels = [
 
 [[package]]
 name = "legendaria"
-version = "0.1.0.dev2"
+version = "0.1.0.dev3"
 source = { editable = "." }
 dependencies = [
     { name = "faster-whisper" },