Adding note on ffmpeg + fix for faster whisper on macOS

README.md

@@ -22,6 +22,9 @@ OpenAI's [Whisper](https://github.com/openai/whisper).
 
 **PyPI**:
 
+Install [ffmpeg](https://www.ffmpeg.org/download.html)
+
+Install Buzz
 ```shell
 pip install buzz-captions
 python -m buzz

buzz/transcriber/recording_transcriber.py

@@ -8,9 +8,9 @@
 from typing import Optional
 from platformdirs import user_cache_dir
 
-import torch
 import numpy as np
 import sounddevice
+from torch import cuda, version
 from sounddevice import PortAudioError
 from openai import OpenAI
 from PyQt6.QtCore import QObject, pyqtSignal
@@ -62,11 +62,11 @@ def start(self):
         model_path = self.model_path
         keep_samples = int(0.15 * self.sample_rate)
 
-        if torch.cuda.is_available():
-            logging.debug(f"CUDA version detected: {torch.version.cuda}")
+        if cuda.is_available():
+            logging.debug(f"CUDA version detected: {version.cuda}")
 
         if self.transcription_options.model.model_type == ModelType.WHISPER:
-            device = "cuda" if torch.cuda.is_available() else "cpu"
+            device = "cuda" if cuda.is_available() else "cpu"
             model = whisper.load_model(model_path, device=device)
         elif self.transcription_options.model.model_type == ModelType.WHISPER_CPP:
             model = WhisperCpp(model_path)
@@ -79,7 +79,7 @@ def start(self):
                 logging.debug("CUDA GPUs are currently no supported on Running on Windows, using CPU")
                 device = "cpu"
 
-            if torch.cuda.is_available() and torch.version.cuda < "12":
+            if cuda.is_available() and version.cuda < "12":
                 logging.debug("Unsupported CUDA version (<12), using CPU")
                 device = "cpu"
 

buzz/transcriber/whisper_file_transcriber.py

@@ -5,8 +5,8 @@
 import re
 import os
 import sys
-import torch
 import platform
+from torch import cuda, version
 from platformdirs import user_cache_dir
 from multiprocessing.connection import Connection
 from threading import Thread
@@ -52,8 +52,8 @@ def transcribe(self) -> List[Segment]:
             "Starting whisper file transcription, task = %s", self.transcription_task
         )
 
-        if torch.cuda.is_available():
-            logging.debug(f"CUDA version detected: {torch.version.cuda}")
+        if cuda.is_available():
+            logging.debug(f"CUDA version detected: {version.cuda}")
 
         recv_pipe, send_pipe = multiprocessing.Pipe(duplex=False)
 
@@ -149,7 +149,7 @@ def transcribe_faster_whisper(cls, task: FileTranscriptionTask) -> List[Segment]
             logging.debug("CUDA GPUs are currently no supported on Running on Windows, using CPU")
             device = "cpu"
 
-        if torch.cuda.is_available() and torch.version.cuda < "12":
+        if cuda.is_available() and version.cuda < "12":
             logging.debug("Unsupported CUDA version (<12), using CPU")
             device = "cpu"
 
@@ -195,7 +195,7 @@ def transcribe_faster_whisper(cls, task: FileTranscriptionTask) -> List[Segment]
 
     @classmethod
     def transcribe_openai_whisper(cls, task: FileTranscriptionTask) -> List[Segment]:
-        device = "cuda" if torch.cuda.is_available() else "cpu"
+        device = "cuda" if cuda.is_available() else "cpu"
         model = whisper.load_model(task.model_path, device=device)
 
         if task.transcription_options.word_level_timings:

buzz/transformers_whisper.py

@@ -1,8 +1,9 @@
 import os
 import sys
 import numpy as np
-import torch
 import requests
+from torch import from_numpy, cuda, float16, float32
+
 from typing import Optional, Union
 from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
 from transformers.pipelines import AutomaticSpeechRecognitionPipeline
@@ -83,7 +84,7 @@ def preprocess(self, inputs, chunk_length_s=0, stride_length_s=None):
                     )
 
                 inputs = F.resample(
-                    torch.from_numpy(inputs), in_sampling_rate, self.feature_extractor.sampling_rate
+                    from_numpy(inputs), in_sampling_rate, self.feature_extractor.sampling_rate
                 ).numpy()
                 ratio = self.feature_extractor.sampling_rate / in_sampling_rate
             else:
@@ -161,8 +162,8 @@ def transcribe(
         language: str,
         task: str,
     ):
-        device = "cuda" if torch.cuda.is_available() else "cpu"
-        torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+        device = "cuda" if cuda.is_available() else "cpu"
+        torch_dtype = float16 if cuda.is_available() else float32
 
         safetensors_path = os.path.join(self.model_id, "model.safetensors")
         use_safetensors = os.path.exists(safetensors_path)