easyaligner 0.2.3__tar.gz → 0.3.1__tar.gz

{easyaligner-0.2.3/src/easyaligner.egg-info → easyaligner-0.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: easyaligner
-Version: 0.2.3
+Version: 0.3.1
 Summary: Forced alignment pipeline designed for efficiency and ease of use.
 Author: Faton Rekathati
 Project-URL: Repository, https://github.com/kb-labb/easyaligner
@@ -8,12 +8,12 @@ Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: transformers>=4.45.0
-Requires-Dist: torch!=2.9.*,>=2.7.0
-Requires-Dist: torchaudio!=2.9.*,>=2.7.0
+Requires-Dist: torch<2.9,>=2.7.0
+Requires-Dist: torchaudio<2.9,>=2.7.0
 Requires-Dist: tqdm>=4.66.1
 Requires-Dist: soundfile>=0.12.1
 Requires-Dist: nltk>=3.8.2
-Requires-Dist: pyannote-audio>=3.3.1
+Requires-Dist: pyannote-audio<4.0.4,>=3.3.1
 Requires-Dist: silero-vad~=6.0
 Requires-Dist: msgspec
 Requires-Dist: rapidfuzz

{easyaligner-0.2.3 → easyaligner-0.3.1}/pyproject.toml

@@ -3,7 +3,7 @@ requires = ["setuptools>=67.0.0"]
 build-backend = "setuptools.build_meta"
 
 [project]
-version = "0.2.3"
+version = "0.3.1"
 name = "easyaligner"
 requires-python = ">= 3.10"
 description = "Forced alignment pipeline designed for efficiency and ease of use."
@@ -12,12 +12,12 @@ authors = [{ name = "Faton Rekathati" }]
 
 dependencies = [
   "transformers>=4.45.0",
-  "torch>=2.7.0,!=2.9.*",
-  "torchaudio>=2.7.0,!=2.9.*",
+  "torch>=2.7.0,<2.9",
+  "torchaudio>=2.7.0,<2.9",
   "tqdm>=4.66.1",
   "soundfile>=0.12.1",
   "nltk>=3.8.2",
-  "pyannote-audio>=3.3.1",
+  "pyannote-audio>=3.3.1,<4.0.4",
   "silero-vad~=6.0",
   "msgspec",
   "rapidfuzz"

{easyaligner-0.2.3 → easyaligner-0.3.1}/src/easyaligner/alignment/utils.py

@@ -231,6 +231,11 @@ def segment_speech_probs(probs_list: list[np.ndarray], speech_ids: list[str] | l
     np.ndarray
         Probabilities for the speech segment.
     """
+    # Nothing to segment (e.g. a file where VAD detected no speech). Yield nothing
+    # so callers iterate over an empty result instead of hitting np.concatenate([]).
+    if not probs_list:
+        return
+
     # Count the number of chunks per speech id
     speech_chunk_counts = [
         (key, sum(1 for i in group)) for key, group in itertools.groupby(speech_ids)

{easyaligner-0.2.3 → easyaligner-0.3.1}/src/easyaligner/vad/pyannote.py

@@ -13,7 +13,7 @@ from pyannote.core import Annotation, Segment, SlidingWindowFeature
 from tqdm import tqdm
 
 from easyaligner.data.datamodel import AudioMetadata, SpeechSegment
-from easyaligner.vad.utils import encode_vad_segments
+from easyaligner.vad.utils import drop_empty_speeches, encode_vad_segments
 
 """
 This file contains modified functions from WhisperX (BSD-4-Clause License).
@@ -431,7 +431,7 @@ def run_vad_pipeline(metadata: AudioMetadata, model, audio, sample_rate=16000, c
         # Run VAD on entire audio
         vad_segments = model(
             {
-                "waveform": torch.tensor(audio).unsqueeze(0).to(torch.float32),
+                "waveform": torch.as_tensor(audio).unsqueeze(0).to(torch.float32),
                 "sample_rate": sample_rate,
             }
         )
@@ -439,12 +439,15 @@ def run_vad_pipeline(metadata: AudioMetadata, model, audio, sample_rate=16000, c
         vad_segments = merge_chunks(vad_segments, chunk_size=chunk_size)
         segments = encode_vad_segments(vad_segments)
 
+        # Create a single SpeechSegment based on where speech was detected.
+        # An empty `speeches` list signals a file with no detected speech.
         metadata.speeches = []
-        metadata.speeches.append(
-            SpeechSegment(
-                start=segments[0].start, end=segments[-1].end, text=None, chunks=segments
+        if segments:
+            metadata.speeches.append(
+                SpeechSegment(
+                    start=segments[0].start, end=segments[-1].end, text=None, chunks=segments
+                )
             )
-        )
     else:
         # Run VAD on each speech segment
         for speech in tqdm(metadata.speeches, desc="Running VAD on speeches"):
@@ -455,7 +458,7 @@ def run_vad_pipeline(metadata: AudioMetadata, model, audio, sample_rate=16000, c
 
             vad_segments = model(
                 {
-                    "waveform": torch.tensor(speech_audio).unsqueeze(0).to(torch.float32),
+                    "waveform": torch.as_tensor(speech_audio).unsqueeze(0).to(torch.float32),
                     "sample_rate": sample_rate,
                 }
             )
@@ -472,11 +475,11 @@ def run_vad_pipeline(metadata: AudioMetadata, model, audio, sample_rate=16000, c
             ]
             segments = encode_vad_segments(vad_segments)
 
-            if speech.duration is None:
+            if speech.duration is None and segments:
                 speech.start = segments[0].start
                 speech.end = segments[-1].end
                 speech.calculate_duration()
 
             speech.chunks = segments  # In place update of chunks in metadata
 
-    return metadata
+    return drop_empty_speeches(metadata)

{easyaligner-0.2.3 → easyaligner-0.3.1}/src/easyaligner/vad/silero.py

@@ -3,7 +3,7 @@ from silero_vad import get_speech_timestamps, load_silero_vad
 from tqdm import tqdm
 
 from easyaligner.data.datamodel import AudioMetadata, SpeechSegment
-from easyaligner.vad.utils import encode_vad_segments
+from easyaligner.vad.utils import drop_empty_speeches, encode_vad_segments
 
 
 def load_vad_model(onnx=False, opset_version=16):
@@ -42,6 +42,9 @@ def merge_chunks(segments, chunk_size=30):
         List of merged chunks, where each chunk is a dictionary with
         "start", "end", and "segments" keys.
     """
+    if not segments:
+        return []
+
     current_start = segments[0]["start"]
     current_end = segments[0]["end"]
     merged_segments = []
@@ -103,17 +106,22 @@ def run_vad_pipeline(
         vad_segments = merge_chunks(vad_segments, chunk_size=chunk_size)
         segments = encode_vad_segments(vad_segments)
 
-        # Create a single SpeechSegment based on where speech was detected
+        # Create a single SpeechSegment based on where speech was detected.
+        # An empty `speeches` list signals a file with no detected speech.
         metadata.speeches = []
-        metadata.speeches.append(
-            SpeechSegment(
-                start=segments[0].start, end=segments[-1].end, text=None, chunks=segments
+        if segments:
+            metadata.speeches.append(
+                SpeechSegment(
+                    start=segments[0].start, end=segments[-1].end, text=None, chunks=segments
+                )
             )
-        )
     else:
         # Run VAD on each speech segment
         for speech in tqdm(metadata.speeches, desc="Running VAD on speeches"):
-            speech_audio = audio[int(speech.start * sample_rate) : int(speech.end * sample_rate)]
+            start = int(speech.start * sample_rate) if speech.start is not None else None
+            end = int(speech.end * sample_rate) if speech.end is not None else None
+            # Note: Using `None` as a slicing parameter is the same as omitting it
+            speech_audio = audio[start:end]
             vad_segments = get_speech_timestamps(
                 speech_audio,
                 model,
@@ -122,15 +130,22 @@ def run_vad_pipeline(
             )
             vad_segments = merge_chunks(vad_segments, chunk_size=chunk_size)
             # Add speech.start offset to each segment
+            offset = speech.start if speech.start is not None else 0
             vad_segments = [
                 {
-                    "start": seg["start"] + speech.start,
-                    "end": seg["end"] + speech.start,
+                    "start": seg["start"] + offset,
+                    "end": seg["end"] + offset,
                     "segments": seg["segments"],
                 }
                 for seg in vad_segments
             ]
             segments = encode_vad_segments(vad_segments)
+
+            if speech.duration is None and segments:
+                speech.start = segments[0].start
+                speech.end = segments[-1].end
+                speech.calculate_duration()
+
             speech.chunks = segments
 
-    return metadata
+    return drop_empty_speeches(metadata)

{easyaligner-0.2.3 → easyaligner-0.3.1}/src/easyaligner/vad/utils.py

@@ -124,6 +124,35 @@ def seconds_to_frames(seconds, sr=16000):
     return int(seconds * sr)
 
 
+def drop_empty_speeches(metadata: AudioMetadata) -> AudioMetadata:
+    """
+    Remove speeches where VAD detected no speech (i.e. speeches without chunks).
+
+    Downstream pipeline stages (emissions extraction, alignment) assume every
+    speech has at least one VAD chunk. An empty `speeches` list signals a file
+    with no detected speech.
+
+    Parameters
+    ----------
+    metadata : AudioMetadata
+        The metadata object to filter after running VAD.
+
+    Returns
+    -------
+    AudioMetadata
+        The metadata object with chunkless speeches removed.
+    """
+    speeches = [speech for speech in metadata.speeches if speech.chunks]
+    num_dropped = len(metadata.speeches) - len(speeches)
+    if num_dropped > 0:
+        logger.warning(
+            f"VAD detected no speech in {num_dropped} speech segment(s) of "
+            f"{metadata.audio_path}. Dropping them from the metadata."
+        )
+    metadata.speeches = speeches
+    return metadata
+
+
 def encode_vad_segments(vad_segments):
     """
     Encode VAD segments into a list of AudioChunk objects.

{easyaligner-0.2.3 → easyaligner-0.3.1/src/easyaligner.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: easyaligner
-Version: 0.2.3
+Version: 0.3.1
 Summary: Forced alignment pipeline designed for efficiency and ease of use.
 Author: Faton Rekathati
 Project-URL: Repository, https://github.com/kb-labb/easyaligner
@@ -8,12 +8,12 @@ Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: transformers>=4.45.0
-Requires-Dist: torch!=2.9.*,>=2.7.0
-Requires-Dist: torchaudio!=2.9.*,>=2.7.0
+Requires-Dist: torch<2.9,>=2.7.0
+Requires-Dist: torchaudio<2.9,>=2.7.0
 Requires-Dist: tqdm>=4.66.1
 Requires-Dist: soundfile>=0.12.1
 Requires-Dist: nltk>=3.8.2
-Requires-Dist: pyannote-audio>=3.3.1
+Requires-Dist: pyannote-audio<4.0.4,>=3.3.1
 Requires-Dist: silero-vad~=6.0
 Requires-Dist: msgspec
 Requires-Dist: rapidfuzz

{easyaligner-0.2.3 → easyaligner-0.3.1}/src/easyaligner.egg-info/requires.txt

@@ -1,10 +1,10 @@
 transformers>=4.45.0
-torch!=2.9.*,>=2.7.0
-torchaudio!=2.9.*,>=2.7.0
+torch<2.9,>=2.7.0
+torchaudio<2.9,>=2.7.0
 tqdm>=4.66.1
 soundfile>=0.12.1
 nltk>=3.8.2
-pyannote-audio>=3.3.1
+pyannote-audio<4.0.4,>=3.3.1
 silero-vad~=6.0
 msgspec
 rapidfuzz