Extract audio from video with FFmpeg: AAC, MP3, WAV, transcription-ready

When to use this

You extract audio when you need the audio independently of the video — podcast extraction from video recordings, ML pipelines (Whisper, Deepgram, AWS Transcribe expect audio-only inputs), audio-only delivery, or building dataset preprocessing. Three common targets: stream-copy AAC (when the original audio is already AAC and you just need a separate file), high-quality MP3 (broader compatibility), and mono 16kHz WAV (the standard input for transcription services).

Command variants

ffmpeg -i input.mp4 \
  -vn -c:a copy \
  output.aac

ffmpeg -i input.mp4 \
  -vn -c:a libmp3lame -q:a 2 \
  output.mp3

ffmpeg -i input.mp4 \
  -vn -ar 16000 -ac 1 -c:a pcm_s16le \
  output.wav

What each parameter does

• -vn

  No video. Tells FFmpeg to drop the video stream entirely and only process audio.

• -c:a copy

  Stream-copy audio without re-encoding. Fastest, preserves original quality exactly. Works only when output container supports the source codec.

• -q:a 2

  libmp3lame VBR quality. Range 0-9; 2 ≈ 190 kbps, 4 ≈ 165 kbps. Lower number = higher quality + larger file.

• -ar 16000

  Audio sample rate 16kHz. Standard for transcription; speech doesn't need higher rates and lower rates reduce file size.

• -ac 1

  Mono (1 channel). Transcription services expect mono; stereo confuses some systems.

• -c:a pcm_s16le

  Uncompressed 16-bit PCM. The format ML transcription tools expect for highest accuracy.

What this outputs

An audio-only file. Stream-copy produces a file with identical audio bytes to the source. MP3 conversion produces a smaller file at slight quality loss. WAV-for-transcription produces a larger file but the format ML tools expect.

Pitfalls

1. Stream-copy fails when the output container doesn't support the source codec. AAC into .mp3 won't work; convert with libmp3lame instead.
2. For transcription pipelines, mono is non-negotiable. Stereo audio inputs to Whisper/Deepgram return errors or unpredictable results. Always -ac 1 for transcription.
3. Higher sample rates than 16kHz waste storage for transcription without improving accuracy. Most speech recognition models downsample to 16kHz internally.
4. MP3 VBR mode (-q:a) and CBR mode (-b:a) produce different file sizes for the same perceived quality. VBR is more efficient; CBR is more predictable.
5. Some video files have multiple audio tracks (commentary, language alternatives). Without explicit -map, FFmpeg picks the first one. Use -map 0:a:0 / 0:a:1 to choose specific tracks.

At production scale

Audio extraction is I/O-bound, not compute-bound. Stream-copy at petabyte scale is essentially free of compute cost; MP3 conversion adds modest overhead. For transcription pipelines processing millions of hours, batch-extracting WAV upstream of the transcription service saves significant retry-on-format-error overhead. Many ML pipelines benefit from a dedicated extraction stage that produces a transcription-ready format once, regardless of how the audio is later consumed.

MpegFlow models audio extraction as a parallel DAG stage that can run alongside video transcoding. The same source file feeds the video pipeline and an audio-extraction pipeline simultaneously, which is more efficient than sequential extraction-then-transcoding.

• FFmpeg
• audio
• extraction
• transcription
• audio-operations

• Recipe
  Loudness Normalize Audio
• Recipe
  Concatenate Video Files