easytranscriber 0.2.3__tar.gz → 0.2.4__tar.gz

{easytranscriber-0.2.3/src/easytranscriber.egg-info → easytranscriber-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: easytranscriber
-Version: 0.2.3
+Version: 0.2.4
 Summary: Speech recognition with accurate word-level timestamps.
 Author: Faton Rekathati
 Project-URL: Repository, https://github.com/kb-labb/easytranscriber

{easytranscriber-0.2.3 → easytranscriber-0.2.4}/pyproject.toml

@@ -3,7 +3,7 @@ requires = ["setuptools>=67.0.0"]
 build-backend = "setuptools.build_meta"
 
 [project]
-version = "0.2.3"
+version = "0.2.4"
 name = "easytranscriber"
 requires-python = ">= 3.10"
 description = "Speech recognition with accurate word-level timestamps."

{easytranscriber-0.2.3 → easytranscriber-0.2.4}/src/easytranscriber/asr/cohere.py

@@ -127,9 +127,11 @@ def transcribe(
                 transcription = processor.batch_decode(outputs, skip_special_tokens=True)
                 transcription_texts.extend(transcription)
 
-        for i, speech in enumerate(metadata.speeches):
-            for j, chunk in enumerate(speech.chunks):
-                chunk.text = transcription_texts[j].strip()
+        global_chunk_idx = 0
+        for speech in metadata.speeches:
+            for chunk in speech.chunks:
+                chunk.text = transcription_texts[global_chunk_idx].strip()
+                global_chunk_idx += 1
 
         output_path = Path(output_dir) / Path(metadata.audio_path).with_suffix(".json")
         output_path.parent.mkdir(parents=True, exist_ok=True)

{easytranscriber-0.2.3 → easytranscriber-0.2.4}/src/easytranscriber/asr/ct2.py

@@ -11,10 +11,11 @@ from pathlib import Path
 import ctranslate2
 import torch
 from easyaligner.utils import save_metadata_json
-from easytranscriber.data.collators import transcribe_collate_fn
 from tqdm import tqdm
 from transformers import WhisperProcessor
 
+from easytranscriber.data.collators import transcribe_collate_fn
+
 logger = logging.getLogger(__name__)
 
 
@@ -90,7 +91,7 @@ def transcribe(
             batch_size=batch_size,
             num_workers=num_workers,
             prefetch_factor=prefetch_factor,
-            pin_memory=True,
+            pin_memory=False,
             collate_fn=transcribe_collate_fn,
         )
 
@@ -198,7 +199,7 @@ def lang_detect_only(
             batch_size=batch_size,
             num_workers=num_workers,
             prefetch_factor=prefetch_factor,
-            pin_memory=True,
+            pin_memory=False,
             collate_fn=transcribe_collate_fn,
         )
 

{easytranscriber-0.2.3 → easytranscriber-0.2.4}/src/easytranscriber/asr/hf.py

@@ -76,7 +76,7 @@ def transcribe(
 
         for batch in feature_dataloader:
             with torch.inference_mode():
-                batch = batch["features"].to(device).half()
+                batch = batch["features"].to(device=device, dtype=model.dtype)
                 predicted_ids = model.generate(
                     batch,
                     return_dict_in_generate=True,

{easytranscriber-0.2.3 → easytranscriber-0.2.4}/src/easytranscriber/pipelines.py

@@ -230,6 +230,9 @@ def pipeline(
     # TODO: Support msgpack throughout the pipeline
     json_paths = [Path(p).with_suffix(".json") for p in audio_paths]
 
+    # fp16 is a GPU optimization; most fp16 ops are unimplemented or slow on CPU.
+    torch_dtype = torch.float16 if "cuda" in str(device) else torch.float32
+
     if streaming:
         DatasetClass = StreamingAudioFileDataset
     else:
@@ -290,7 +293,7 @@ def pipeline(
         logger.info(f"Loading Cohere ASR model from {transcription_model}...")
         model = (
             CohereAsrForConditionalGeneration.from_pretrained(
-                transcription_model, torch_dtype=torch.float16, cache_dir=cache_dir
+                transcription_model, torch_dtype=torch_dtype, cache_dir=cache_dir
             )
             .to(device)
             .eval()
@@ -320,7 +323,7 @@ def pipeline(
         else:
             logger.info(f"Loading Hugging Face model from {transcription_model}...")
             model = WhisperForConditionalGeneration.from_pretrained(
-                transcription_model, torch_dtype=torch.float16, cache_dir=cache_dir
+                transcription_model, torch_dtype=torch_dtype, cache_dir=cache_dir
             ).to(device)
 
         processor = WhisperProcessor.from_pretrained(transcription_model, cache_dir=cache_dir)
@@ -368,7 +371,11 @@ def pipeline(
         json_paths=[str(Path(output_transcriptions_dir) / p) for p in json_paths]
     )
 
-    model = AutoModelForCTC.from_pretrained(emissions_model, cache_dir=cache_dir).to("cuda").half()
+    model = (
+        AutoModelForCTC.from_pretrained(emissions_model, cache_dir=cache_dir)
+        .to(device)
+        .to(torch_dtype)
+    )
     processor = Wav2Vec2Processor.from_pretrained(emissions_model, cache_dir=cache_dir)
 
     if blank_id is None:
@@ -394,6 +401,7 @@ def pipeline(
         save_emissions=save_emissions,
         return_emissions=False,
         output_dir=output_emissions_dir,
+        device=device,
     )
 
     # Step 4: Perform Alignment

{easytranscriber-0.2.3 → easytranscriber-0.2.4/src/easytranscriber.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: easytranscriber
-Version: 0.2.3
+Version: 0.2.4
 Summary: Speech recognition with accurate word-level timestamps.
 Author: Faton Rekathati
 Project-URL: Repository, https://github.com/kb-labb/easytranscriber