Merge ae7ea9f4b6 into b343241253

Fixes issue where `np.frombuffer()` would crash on large audio files due to excessive memory usage or misaligned buffer sizes.

.github/workflows/build-and-release.yml

@@ -17,6 +17,9 @@ jobs:
           version: "0.5.14"
           python-version: "3.9"
 
+      - name: Check if lockfile is up to date
+        run: uv lock --check
+
       - name: Build package
         run: uv build
 

.github/workflows/python-compatibility.yml

@@ -23,6 +23,9 @@ jobs:
           version: "0.5.14"
           python-version: ${{ matrix.python-version }}
 
+      - name: Check if lockfile is up to date
+        run: uv lock --check
+
       - name: Install the project
         run: uv sync --all-extras
 

README.md

@@ -97,6 +97,25 @@ uv sync --all-extras --dev
 
 You may also need to install ffmpeg, rust etc. Follow openAI instructions here https://github.com/openai/whisper#setup.
 
+### Common Issues & Troubleshooting 🔧
+
+#### libcudnn Dependencies (GPU Users)
+
+If you're using WhisperX with GPU support and encounter errors like:
+
+- `Could not load library libcudnn_ops_infer.so.8`
+- `Unable to load any of {libcudnn_cnn.so.9.1.0, libcudnn_cnn.so.9.1, libcudnn_cnn.so.9, libcudnn_cnn.so}`
+- `libcudnn_ops_infer.so.8: cannot open shared object file: No such file or directory`
+
+This means your system is missing the CUDA Deep Neural Network library (cuDNN). This library is needed for GPU acceleration but isn't always installed by default.
+
+**Install cuDNN (example for apt based systems):**
+
+```bash
+sudo apt update
+sudo apt install libcudnn8 libcudnn8-dev -y
+```
+
 ### Speaker Diarization
 
 To **enable Speaker Diarization**, include your Hugging Face access token (read) that you can generate from [Here](https://huggingface.co/settings/tokens) after the `--hf_token` argument and accept the user agreement for the following models: [Segmentation](https://huggingface.co/pyannote/segmentation-3.0) and [Speaker-Diarization-3.1](https://huggingface.co/pyannote/speaker-diarization-3.1) (if you choose to use Speaker-Diarization 2.x, follow requirements [here](https://huggingface.co/pyannote/speaker-diarization) instead.)

uv.lock

@@ -2787,7 +2787,7 @@ wheels = [
 
 [[package]]
 name = "whisperx"
-version = "3.3.3"
+version = "3.3.4"
 source = { editable = "." }
 dependencies = [
     { name = "ctranslate2" },

whisperx/__main__.py

@@ -43,6 +43,7 @@ def cli():
     parser.add_argument("--diarize", action="store_true", help="Apply diarization to assign speaker labels to each segment/word")
     parser.add_argument("--min_speakers", default=None, type=int, help="Minimum number of speakers to in audio file")
     parser.add_argument("--max_speakers", default=None, type=int, help="Maximum number of speakers to in audio file")
+    parser.add_argument("--diarize_model", default="pyannote/speaker-diarization-3.1", type=str, help="Name of the speaker diarization model to use")
 
     parser.add_argument("--temperature", type=float, default=0, help="temperature to use for sampling")
     parser.add_argument("--best_of", type=optional_int, default=5, help="number of candidates when sampling with non-zero temperature")

whisperx/audio.py

@@ -58,11 +58,33 @@ def load_audio(file: str, sr: int = SAMPLE_RATE) -> np.ndarray:
             str(sr),
             "-",
         ]
-        out = subprocess.run(cmd, capture_output=True, check=True).stdout
-    except subprocess.CalledProcessError as e:
-        raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e
 
-    return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
+        process = subprocess.Popen(
+            cmd,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            bufsize=10 * 1024 * 1024
+        )
+
+        out = bytearray()
+        while True:
+            chunk = process.stdout.read(1024 * 1024)
+            if not chunk:
+                break
+            out.extend(chunk)
+
+        stderr_output = process.stderr.read()
+        return_code = process.wait()
+
+        if return_code != 0:
+            raise RuntimeError(f"FFmpeg process failed with error: {stderr_output.decode()}")
+
+        if len(out) % 2 != 0:
+            raise ValueError("Audio buffer size is not aligned to int16.")
+        
+        return np.frombuffer(out, np.int16).astype(np.float32) / 32768.0
+    except Exception as e:
+        raise RuntimeError(f"Error loading audio file {file}: {str(e)}")
 
 
 def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):

whisperx/diarize.py

@@ -11,13 +11,14 @@ from whisperx.types import TranscriptionResult, AlignedTranscriptionResult
 class DiarizationPipeline:
     def __init__(
         self,
-        model_name="pyannote/speaker-diarization-3.1",
+        model_name=None,
         use_auth_token=None,
         device: Optional[Union[str, torch.device]] = "cpu",
     ):
         if isinstance(device, str):
             device = torch.device(device)
-        self.model = Pipeline.from_pretrained(model_name, use_auth_token=use_auth_token).to(device)
+        model_config = model_name or "pyannote/speaker-diarization-3.1"
+        self.model = Pipeline.from_pretrained(model_config, use_auth_token=use_auth_token).to(device)
 
     def __call__(
         self,

whisperx/transcribe.py

@@ -57,6 +57,7 @@ def transcribe_task(args: dict, parser: argparse.ArgumentParser):
     diarize: bool = args.pop("diarize")
     min_speakers: int = args.pop("min_speakers")
     max_speakers: int = args.pop("max_speakers")
+    diarize_model_name: str = args.pop("diarize_model")
     print_progress: bool = args.pop("print_progress")
 
     if args["language"] is not None:
@@ -204,8 +205,9 @@ def transcribe_task(args: dict, parser: argparse.ArgumentParser):
             )
         tmp_results = results
         print(">>Performing diarization...")
+        print(">>Using model:", diarize_model_name)
         results = []
-        diarize_model = DiarizationPipeline(use_auth_token=hf_token, device=device)
+        diarize_model = DiarizationPipeline(model_name=diarize_model_name, use_auth_token=hf_token, device=device)
         for result, input_audio_path in tmp_results:
             diarize_segments = diarize_model(
                 input_audio_path, min_speakers=min_speakers, max_speakers=max_speakers