videopython 0.27.0__tar.gz → 0.27.2__tar.gz

{videopython-0.27.0 → videopython-0.27.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: videopython
-Version: 0.27.0
+Version: 0.27.2
 Summary: Minimal video generation and processing library.
 Project-URL: Homepage, https://videopython.com
 Project-URL: Repository, https://github.com/bartwojtowicz/videopython/

{videopython-0.27.0 → videopython-0.27.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "videopython"
-version = "0.27.0"
+version = "0.27.2"
 description = "Minimal video generation and processing library."
 authors = [
     { name = "Bartosz Wójtowicz", email = "bartoszwojtowicz@outlook.com" },

{videopython-0.27.0 → videopython-0.27.2}/src/videopython/ai/dubbing/dubber.py

@@ -28,18 +28,32 @@ class VideoDubber:
         whisper_model: Whisper model size used for transcription. Larger models
             give better accuracy at the cost of VRAM and latency. One of
             ``tiny``, ``base``, ``small``, ``medium``, ``large``, ``turbo``.
-            Default ``small``.
+            Default ``turbo``.
+        condition_on_previous_text: Forwarded to ``AudioToText``. Defaults to
+            ``False`` (Whisper's own default is ``True``). With conditioning on,
+            a single hallucinated filler phrase cascades through the rest of
+            the file. See ``AudioToText`` for the full rationale.
+        no_speech_threshold: Forwarded to ``AudioToText``. Whisper's no-speech
+            gate; raise to drop more low-confidence windows.
+        logprob_threshold: Forwarded to ``AudioToText``. Whisper's average
+            log-probability gate.
     """
 
     def __init__(
         self,
         device: str | None = None,
         low_memory: bool = False,
-        whisper_model: WhisperModel = "small",
+        whisper_model: WhisperModel = "turbo",
+        condition_on_previous_text: bool = False,
+        no_speech_threshold: float = 0.6,
+        logprob_threshold: float | None = -1.0,
     ):
         self.device = device
         self.low_memory = low_memory
         self.whisper_model = whisper_model
+        self.condition_on_previous_text = condition_on_previous_text
+        self.no_speech_threshold = no_speech_threshold
+        self.logprob_threshold = logprob_threshold
         self._local_pipeline: Any = None
         requested = device.lower() if isinstance(device, str) else "auto"
         logger.info(
@@ -56,6 +70,9 @@ class VideoDubber:
             device=self.device,
             low_memory=self.low_memory,
             whisper_model=self.whisper_model,
+            condition_on_previous_text=self.condition_on_previous_text,
+            no_speech_threshold=self.no_speech_threshold,
+            logprob_threshold=self.logprob_threshold,
         )
 
     def dub(

{videopython-0.27.0 → videopython-0.27.2}/src/videopython/ai/dubbing/pipeline.py

@@ -60,11 +60,17 @@ class LocalDubbingPipeline:
         self,
         device: str | None = None,
         low_memory: bool = False,
-        whisper_model: WhisperModel = "small",
+        whisper_model: WhisperModel = "turbo",
+        condition_on_previous_text: bool = False,
+        no_speech_threshold: float = 0.6,
+        logprob_threshold: float | None = -1.0,
     ):
         self.device = device
         self.low_memory = low_memory
         self.whisper_model = whisper_model
+        self.condition_on_previous_text = condition_on_previous_text
+        self.no_speech_threshold = no_speech_threshold
+        self.logprob_threshold = logprob_threshold
         requested = device.lower() if isinstance(device, str) else "auto"
         logger.info(
             "LocalDubbingPipeline initialized with device=%s low_memory=%s whisper_model=%s",
@@ -106,6 +112,9 @@ class LocalDubbingPipeline:
             model_name=self.whisper_model,
             device=self.device,
             enable_diarization=enable_diarization,
+            condition_on_previous_text=self.condition_on_previous_text,
+            no_speech_threshold=self.no_speech_threshold,
+            logprob_threshold=self.logprob_threshold,
         )
 
     def _init_translator(self) -> None:

{videopython-0.27.0 → videopython-0.27.2}/src/videopython/ai/understanding/audio.py

@@ -15,21 +15,45 @@ class AudioToText:
     """Transcription service for audio and video using local Whisper models.
 
     Uses openai-whisper for transcription (with word-level timestamps) and
-    pyannote-audio for optional speaker diarization.
+    pyannote-audio for optional speaker diarization. By default, Silero VAD
+    runs before Whisper to gate language detection on a 30s window built from
+    voiced regions only — fixes Whisper's tendency to lock onto the wrong
+    language when the file opens with silence, music, or non-vocal credits.
+    Disable with ``enable_vad=False`` to reproduce pre-0.27 behaviour.
+
+    Three Whisper decoder kwargs are surfaced for anti-hallucination tuning:
+
+    - ``condition_on_previous_text`` defaults to ``False`` (Whisper's own
+      default is ``True``). With conditioning on, a single hallucinated filler
+      phrase cascades through the rest of the file because each window's
+      decoder is primed by the previous window's decoded text. Turning it off
+      is the most commonly recommended fix for that failure mode; the cost on
+      clean audio is small (slightly less context for ambiguous homophones
+      across sentence boundaries).
+    - ``no_speech_threshold`` and ``logprob_threshold`` are forwarded with
+      Whisper's documented defaults (``0.6`` and ``-1.0``); raising
+      ``no_speech_threshold`` biases toward dropping low-confidence windows
+      instead of emitting filler.
     """
 
     PYANNOTE_DIARIZATION_MODEL = "pyannote/speaker-diarization-community-1"
 
     def __init__(
         self,
-        model_name: Literal["tiny", "base", "small", "medium", "large", "turbo"] = "small",
+        model_name: Literal["tiny", "base", "small", "medium", "large", "turbo"] = "turbo",
         enable_diarization: bool = False,
         enable_vad: bool = True,
+        condition_on_previous_text: bool = False,
+        no_speech_threshold: float = 0.6,
+        logprob_threshold: float | None = -1.0,
         device: str | None = None,
     ):
         self.model_name = model_name
         self.enable_diarization = enable_diarization
         self.enable_vad = enable_vad
+        self.condition_on_previous_text = condition_on_previous_text
+        self.no_speech_threshold = no_speech_threshold
+        self.logprob_threshold = logprob_threshold
         self.device = select_device(device, mps_allowed=False)
         log_device_initialization(
             "AudioToText",
@@ -40,6 +64,16 @@ class AudioToText:
         self._diarization_pipeline: Any = None
         self._vad_model: Any = None
 
+    def _transcribe_kwargs(self, language: str | None) -> dict[str, Any]:
+        """Kwargs threaded into ``whisper.Whisper.transcribe`` from both call sites."""
+        return {
+            "word_timestamps": True,
+            "language": language,
+            "condition_on_previous_text": self.condition_on_previous_text,
+            "no_speech_threshold": self.no_speech_threshold,
+            "logprob_threshold": self.logprob_threshold,
+        }
+
     def _init_local(self) -> None:
         """Initialize local Whisper model."""
         import whisper
@@ -249,7 +283,7 @@ class AudioToText:
             self._init_diarization()
 
         audio_data = audio_mono.data
-        transcription_result = self._model.transcribe(audio=audio_data, word_timestamps=True, language=language)
+        transcription_result = self._model.transcribe(audio=audio_data, **self._transcribe_kwargs(language))
 
         waveform = torch.from_numpy(audio_data.astype(np.float32)).unsqueeze(0)
         diarization_result = self._diarization_pipeline(
@@ -296,7 +330,7 @@ class AudioToText:
         if self.enable_diarization:
             return self._transcribe_with_diarization(audio_mono, language)
 
-        transcription_result = self._model.transcribe(audio=audio_mono.data, word_timestamps=True, language=language)
+        transcription_result = self._model.transcribe(audio=audio_mono.data, **self._transcribe_kwargs(language))
         return self._process_transcription_result(transcription_result)
 
     def transcribe(self, media: Audio | Video) -> Transcription: