使用 Whisper AI 转录时出现打字错误

Question

我想使用 Whisper AI 转录音频文件。 我从一篇文章中学到了https://www.assembleai.com/blog/how-to-run-openais-whisper-speech-recognition-model/ 使用 python 版本 3.8.16 和 pytorch 1.12.1 就可以完成这项工作。

我还安装了最新版本的ffmpeg

我使用 jupyter 笔记本来运行我的代码

import sys
print(sys.version)

import torch
print(torch.__version__)

退货

3.8.16 | packaged by conda-forge | (default, Feb  1 2023, 15:53:35) [MSC v.1929 64 bit (AMD64)]
1.12.1

然后我使用这段代码尝试转录测试音频：

import openai
import whisper

openai.api_key = "my open ai key"

audio_file= open(r"C:\Users\Chris\Python scripts\test_audio.mp3", "rb")

model = whisper.load_model("base")
result = model.transcribe(audio_file)

print(result)

运行此程序时出现以下错误

TypeError                                 Traceback (most recent call last)
Cell In[28], line 9
      6 audio_file= open(r"C:\Users\Chris\Python scripts\test_audio.mp3", "rb")
      8 model = whisper.load_model("base")
----> 9 result = model.transcribe(audio_file)
     11 print(result)

File ~\anaconda3\lib\site-packages\whisper\transcribe.py:121, in transcribe(model, audio, verbose, temperature, compression_ratio_threshold, logprob_threshold, no_speech_threshold, condition_on_previous_text, initial_prompt, word_timestamps, prepend_punctuations, append_punctuations, **decode_options)
    118     decode_options["fp16"] = False
    120 # Pad 30-seconds of silence to the input audio, for slicing
--> 121 mel = log_mel_spectrogram(audio, padding=N_SAMPLES)
    122 content_frames = mel.shape[-1] - N_FRAMES
    124 if decode_options.get("language", None) is None:

File ~\anaconda3\lib\site-packages\whisper\audio.py:131, in log_mel_spectrogram(audio, n_mels, padding, device)
    129     if isinstance(audio, str):
    130         audio = load_audio(audio)
--> 131     audio = torch.from_numpy(audio)
    133 if device is not None:
    134     audio = audio.to(device)

TypeError: expected np.ndarray (got _io.BufferedReader)

为了让它发挥作用，我缺少什么？

Answer 1

你不需要导入 openAI 并为我耳语这个代码工作

    import openai
    
    openai.api_key = api_key
    
    audio_file_path = "./audio.mp3"
    
    with open(audio_file_path, "rb") as audio_file:
        response = openai.Audio.transcribe(model="whisper-1", engine="whisper",response_format="text", file=audio_file)
    
    print(response)

Answer 2

model.transcribe

的第一个参数是

Union[str, np.ndarray, torch.Tensor]

类型——这意味着它需要一个字符串（即文件路径）、一个numpy数组或一个torch张量。尝试直接传递文件路径而不是打开它：

import whisper

audio_file= "C:\Users\Chris\Python scripts\test_audio.mp3"

model = whisper.load_model("base")
result = model.transcribe(audio_file)

print(result)

另外，由于您是离线工作，因此无需设置

openai.api_key

。