如何无需写入本地文件，将YouTube音频流直接转为语音可识别格式？

问题 #

我原本使用以下代码转录YouTube视频的音频流：

from pytube import YouTube
import speech_recognition as sr
from moviepy.editor import *

video = YouTube(url=link)
audio_stream = video.streams.get_by_itag(140)

recognizer = sr.Recognizer()

audio_stream.download(filename="mp4_output.mp4")
audio = AudioFileClip("mp4_output.mp4")
audio.write_audiofile("wav_output.wav")

with sr.AudioFile("./wav_output.wav") as audio_file:
    audio_data = recognizer.record(audio_file, duration=100)
    transcript = recognizer.recognize_sphinx(audio_data=audio_data)

但该流程需要先下载MP4文件再转换为WAV格式，效率较低。我尝试直接将流写入缓冲区处理，代码如下：

from pytube import YouTube
import speech_recognition as sr

video = YouTube(url=link)
audio_stream = video.streams.get_by_itag(140)

buffer = io.BytesIO()
audio_stream.stream_to_buffer(buffer)

recognizer = sr.Recognizer()

with sr.AudioFile(buffer) as audio_file:
    audio_data = recognizer.record(audio_file, duration=100)
    transcript = recognizer.recognize_sphinx(audio_data=audio_data) 

运行后出现报错：audio file could not be read as PCM WAV, AIFF/AIFF-C, or Native FLAC; check if file is corrupted or in another format，请问如何将缓冲区中的内容转换为支持的格式？

解决方案 #

报错的核心原因是：stream_to_buffer写入缓冲区的是MP4封装的音频数据，而speech_recognition的AudioFile仅支持PCM WAV、AIFF/AIFF-C或原生FLAC格式。要在内存中完成格式转换，无需落地文件，可采用以下两种方法：

方法1：使用moviepy处理内存缓冲区 #

通过两个BytesIO缓冲区分别存储MP4源数据和转换后的WAV数据，全程在内存中完成格式转换：

from pytube import YouTube
import speech_recognition as sr
from moviepy.editor import AudioFileClip
import io

# 获取YouTube音频流并写入MP4缓冲区
video = YouTube(url=link)
audio_stream = video.streams.get_by_itag(140)
mp4_buffer = io.BytesIO()
audio_stream.stream_to_buffer(mp4_buffer)
mp4_buffer.seek(0)  # 重置缓冲区指针至起始位置

# 将MP4转换为WAV并写入新缓冲区
wav_buffer = io.BytesIO()
with AudioFileClip(mp4_buffer) as audio_clip:
    audio_clip.write_audiofile(wav_buffer, codec='pcm_s16le')
wav_buffer.seek(0)  # 重置WAV缓冲区指针

# 执行语音识别
recognizer = sr.Recognizer()
with sr.AudioFile(wav_buffer) as audio_file:
    audio_data = recognizer.record(audio_file, duration=100)
    transcript = recognizer.recognize_sphinx(audio_data=audio_data)

print(transcript)

方法2：使用ffmpeg-python直接内存转码（更高效） #

若本地已安装ffmpeg环境，使用ffmpeg-python可实现更高效的内存转码，避免moviepy的额外封装开销：

from pytube import YouTube
import speech_recognition as sr
import io
import ffmpeg

# 获取YouTube音频流的MP4字节数据
video = YouTube(url=link)
audio_stream = video.streams.get_by_itag(140)
mp4_buffer = io.BytesIO()
audio_stream.stream_to_buffer(mp4_buffer)
mp4_buffer.seek(0)

# 用ffmpeg在内存中将MP4转码为WAV
process = (
    ffmpeg
    .input('pipe:', format='mp4')
    .output('pipe:', format='wav', acodec='pcm_s16le')
    .run_async(pipe_stdin=True, pipe_stdout=True)
)
wav_data, _ = process.communicate(input=mp4_buffer.getvalue())
wav_buffer = io.BytesIO(wav_data)
wav_buffer.seek(0)

# 执行语音识别
recognizer = sr.Recognizer()
with sr.AudioFile(wav_buffer) as audio_file:
    audio_data = recognizer.record(audio_file, duration=100)
    transcript = recognizer.recognize_sphinx(audio_data=audio_data)

print(transcript)

关键注意点 #

• 每次写入缓冲区后必须调用seek(0)，否则后续读取操作会从缓冲区末尾开始，导致读取空数据。
• 使用moviepy需确保已安装moviepy包；使用ffmpeg-python需安装ffmpeg-python包，且本地环境配置了ffmpeg可执行文件。

内容的提问来源于stack exchange，提问作者cdt123