pixeltable 0.2.23__tar.gz → 0.2.24__tar.gz

{pixeltable-0.2.23 → pixeltable-0.2.24}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: pixeltable
-Version: 0.2.23
+Version: 0.2.24
 Summary: Pixeltable: The Multimodal AI Data Plane
 Author: Pixeltable, Inc.
 Author-email: contact@pixeltable.com

{pixeltable-0.2.23 → pixeltable-0.2.24}/pixeltable/__version__.py

@@ -1,3 +1,3 @@
 # These version placeholders will be replaced during build.
-__version__ = "0.2.23"
-__version_tuple__ = (0, 2, 23)
+__version__ = "0.2.24"
+__version_tuple__ = (0, 2, 24)

{pixeltable-0.2.23 → pixeltable-0.2.24}/pixeltable/functions/huggingface.py

@@ -426,12 +426,10 @@ def speech2text_for_conditional_generation(
     env.Env.get().require_package('torchaudio')
     env.Env.get().require_package('sentencepiece')
     device = resolve_torch_device('auto', allow_mps=False)  # Doesn't seem to work on 'mps'; use 'cpu' instead
-    import librosa
     import torch
+    import torchaudio  # type: ignore[import-untyped]
     from transformers import Speech2TextForConditionalGeneration, Speech2TextProcessor
 
-    # facebook/s2t-small-librispeech-asr
-    # facebook/s2t-small-mustc-en-fr-st
     model = _lookup_model(model_id, Speech2TextForConditionalGeneration.from_pretrained, device=device)
     processor = _lookup_processor(model_id, Speech2TextProcessor.from_pretrained)
     assert isinstance(processor, Speech2TextProcessor)
@@ -445,12 +443,23 @@ def speech2text_for_conditional_generation(
 
     # Get the model's sampling rate. Default to 16 kHz (the standard) if not in config
     model_sampling_rate = getattr(model.config, 'sampling_rate', 16_000)
-    waveform, sampling_rate = librosa.load(audio, sr=model_sampling_rate, mono=True)
+
+    waveform, sampling_rate = torchaudio.load(audio)
+
+    # Resample to the model's sampling rate, if necessary
+    if sampling_rate != model_sampling_rate:
+        waveform = torchaudio.transforms.Resample(sampling_rate, model_sampling_rate)(waveform)
+
+    # Average the channels to get a single-channel waveform as a 1D tensor (if the original waveform is already
+    # mono, this will simply squeeze the tensor)
+    assert waveform.dim() == 2
+    waveform = torch.mean(waveform, dim=0)
+    assert waveform.dim() == 1
 
     with torch.no_grad():
         inputs = processor(
             waveform,
-            sampling_rate=sampling_rate,
+            sampling_rate=model_sampling_rate,
             return_tensors='pt'
         )
         generated_ids = model.generate(**inputs.to(device), forced_bos_token_id=forced_bos_token_id).to('cpu')

{pixeltable-0.2.23 → pixeltable-0.2.24}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry_dynamic_versioning.backend"
 
 [tool.poetry]
 name = "pixeltable"
-version = "0.2.23"
+version = "0.2.24"
 description = "Pixeltable: The Multimodal AI Data Plane"
 authors = ["Pixeltable, Inc. <contact@pixeltable.com>"]
 readme = "README.md"