Drop ffmpeg-python dependency and call ffmpeg directly. (openai#1242)

* Drop ffmpeg-python dependency and call ffmpeg directly.

The last ffmpeg-python module release was in 2019[1], upstream seem to be
unavailable[2] and the project development seem to have stagnated[3].  As
the features it provide is trivial to replace using the Python native
subprocess module, drop the dependency.

 [1] <URL: https://github.com/kkroening/ffmpeg-python/tags >
 [2] <URL: kkroening/ffmpeg-python#760 >
 [3] <URL: https://openhub.net/p/ffmpeg-python >

* Rewrote to use subprocess.run() instead of subprocess.Popen().

* formatting changes

* formatting update

* isort fix

* Error checking

* isort 🤦🏻

* flake8 fix

* minor spelling changes

---------

Co-authored-by: Jong Wook Kim <jongwook@openai.com>

README.md

@@ -17,9 +17,7 @@ A Transformer sequence-to-sequence model is trained on various speech processing
 
 ## Setup
 
-
-We used Python 3.9.9 and [PyTorch](https://pytorch.org/) 1.10.1 to train and test our models, but the codebase is expected to be compatible with Python 3.8-3.11 and recent PyTorch versions. The codebase also depends on a few Python packages, most notably [OpenAI's tiktoken](https://github.com/openai/tiktoken) for their fast tokenizer implementation and [ffmpeg-python](https://github.com/kkroening/ffmpeg-python) for reading audio files. You can download and install (or update to) the latest release of Whisper with the following command:
-
+We used Python 3.9.9 and [PyTorch](https://pytorch.org/) 1.10.1 to train and test our models, but the codebase is expected to be compatible with Python 3.8-3.11 and recent PyTorch versions. The codebase also depends on a few Python packages, most notably [OpenAI's tiktoken](https://github.com/openai/tiktoken) for their fast tokenizer implementation. You can download and install (or update to) the latest release of Whisper with the following command:
 
     pip install -U openai-whisper
 

requirements.txt

@@ -4,4 +4,3 @@ torch
 tqdm
 more-itertools
 tiktoken==0.3.3
-ffmpeg-python==0.2.0

whisper/audio.py

@@ -1,8 +1,8 @@
 import os
 from functools import lru_cache
+from subprocess import CalledProcessError, run
 from typing import Optional, Union
 
-import ffmpeg
 import numpy as np
 import torch
 import torch.nn.functional as F
@@ -39,15 +39,25 @@ def load_audio(file: str, sr: int = SAMPLE_RATE):
     -------
     A NumPy array containing the audio waveform, in float32 dtype.
     """
+
+    # This launches a subprocess to decode audio while down-mixing
+    # and resampling as necessary.  Requires the ffmpeg CLI in PATH.
+    # fmt: off
+    cmd = [
+        "ffmpeg",
+        "-nostdin",
+        "-threads", "0",
+        "-i", file,
+        "-f", "s16le",
+        "-ac", "1",
+        "-acodec", "pcm_s16le",
+        "-ar", str(sr),
+        "-"
+    ]
+    # fmt: on
     try:
-        # This launches a subprocess to decode audio while down-mixing and resampling as necessary.
-        # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
-        out, _ = (
-            ffmpeg.input(file, threads=0)
-            .output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=sr)
-            .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
-        )
-    except ffmpeg.Error as e:
+        out = run(cmd, capture_output=True, check=True).stdout
+    except 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