Speech-to-text services only process the audio channel. When you upload a video file, the service extracts the audio internally before transcription begins. You are uploading gigabytes of pixel data that the service immediately discards.
Extracting the audio yourself before uploading saves significant time. A 60-minute 1080p video might be 2-4 GB. The audio from that same video is 50-80 MB. Upload speed is the bottleneck for most users, and reducing the file by 97% eliminates that bottleneck.
Which format transcription services prefer
Most transcription services accept multiple formats, but some work better than others:
Whisper (OpenAI's open-source model): Accepts MP3, MP4, M4A, WAV, WEBM. MP3 at 128 kbps is the most efficient choice for spoken word.
Otter.ai: Accepts MP3, M4A, WAV, WMA, MP4. Recommends mono audio at 16 kHz for best accuracy. Higher sample rates work but do not improve transcription quality for speech.
Rev: Accepts most common formats. No specific format preference documented.
Google Cloud Speech-to-Text: Prefers FLAC or WAV for highest accuracy. Accepts MP3 and OGG. Recommends 16 kHz sample rate for telephony audio and 48 kHz for high-fidelity recordings.
AssemblyAI: Accepts any common audio or video format via URL. For uploads, MP3 or M4A are most efficient.
For all of these services, MP3 at 128-192 kbps is a safe universal choice for speech content. Higher bitrates do not improve transcription accuracy because speech occupies a narrow frequency range (roughly 300 Hz to 3400 Hz for intelligibility, up to 8 kHz for natural-sounding speech).
Mono vs. stereo for transcription
Speech transcription services work on mono audio. If your source is stereo (two channels), converting to mono during extraction halves the audio file size with zero impact on transcription accuracy.
Some recordings have the speaker on one channel and audience/ambient noise on the other. For these, extracting only the speaker's channel can improve transcription accuracy by removing background interference.
Sample rate for transcription
Human speech is fully captured at 16 kHz sample rate (which captures frequencies up to 8 kHz, covering the entire speech spectrum). Most video records audio at 44.1 kHz or 48 kHz, which is designed for music reproduction. Downsampling from 48 kHz to 16 kHz during extraction further reduces file size without affecting speech transcription quality.
However, most modern transcription services handle this downsampling internally, so you do not need to do it manually unless file size is a critical concern (for example, uploading over a very slow connection).
Batch extraction for multiple recordings
If you have multiple video recordings to transcribe (a week of meeting recordings, a series of lecture captures), extracting audio from all of them first creates a batch of small files that upload quickly.
Process: Extract audio from each video individually, name the outputs clearly (meeting-2026-01-15.mp3, meeting-2026-01-16.mp3), then upload the batch to your transcription service. The MediaBrew extract audio tool outputs MP3 by default for maximum transcription service compatibility. Processing runs locally so the video footage never reaches an external server.