The short answer: format matters less than people think. Sample rate, bit depth, and whether the file was already lossy on the way in matter a lot more. If your transcripts are coming back garbled, the file extension is rarely the culprit.
We see this question every week from teams setting up a transcription pipeline. Someone reads "WAV is best" on a forum, spends an afternoon batch-converting their library from MP3 to WAV, and gets the same word error rate they had yesterday. The codec wasn't the problem.
Here's what actually moves the needle, and what's just noise.
Does the audio format actually affect transcription accuracy?
Slightly. The real factors are upstream of the format.
Modern speech models like Whisper and faster-whisper resample everything internally to 16 kHz mono before they do anything else. So if you hand them a 96 kHz stereo studio recording, they down-convert it on the spot. The format you saved in isn't what the model hears.
What does matter:
- Was the recording lossy at the source? A phone call captured at 8 kHz has already thrown away the upper half of the speech band. No re-encoding fixes that.
- How aggressive was the compression? A 64 kbps MP3 of a noisy room is going to transcribe worse than a 256 kbps MP3 of the same room. The bitrate decides what's preserved.
- Is the speech intelligible to a human? If you can't make out a word listening at normal volume, the model probably can't either.
We covered the upstream side of this in best practices for audio quality before transcribing. Format is the last step in that chain, not the first.
Which audio format gives the best transcription results?
If you're picking from scratch and disk space isn't a concern, 16-bit WAV at 16 kHz, mono, is the cleanest choice. It's what Whisper expects internally, it's lossless, and there's nothing to second-guess.
In practice, almost anything sensible works just as well:
- WAV (16-bit, 16 kHz mono): the reference, uncompressed
- FLAC: lossless, about half the file size of WAV
- MP3 at 128 kbps or higher: fine for clear speech
- M4A / AAC at 96 kbps or higher: fine, very common output from phones and Zoom
- OGG / Opus at 64 kbps or higher: Opus is engineered for voice and holds up well
The differences between any of these on a clean recording are usually inside the model's own variance. We've run the same 30-minute interview through Whisper as a 256 kbps MP3 and as a 16-bit WAV and gotten transcripts that differ by one or two words out of several thousand.
WAV vs MP3 for transcription: what's the real difference?
WAV is uncompressed PCM. MP3 is lossy compression. WAV files are roughly 10x the size of a comparable MP3.
For transcription specifically:
- WAV is "safer" in the sense that you can re-encode it later without further loss. Every time you transcode an MP3, you lose a little more.
- MP3 at 128 kbps and above preserves the speech band well enough that accuracy barely budges.
- MP3 below 64 kbps starts to do audible damage to consonants and sibilants. That's where word error rate creeps up.
The honest take: if your source is already an MP3 at a reasonable bitrate, don't convert it to WAV before transcribing. You're not recovering anything. You're just making the file ten times bigger.
What sample rate and bit depth should you use?
For speech, 16 kHz, 16-bit is the sweet spot.
The reason is the Nyquist sampling theorem: to capture frequencies up to f Hz, you need to sample at 2f Hz. Human speech has meaningful content up to about 8 kHz (sibilants like "s" and "f" live up there). So 16 kHz sampling captures everything that matters for word recognition.
- 8 kHz (telephone audio) cuts off at 4 kHz. Sibilants and fricatives get muddy. Names with similar-sounding consonants ("Beth" vs "Seth") become coin flips.
- 16 kHz captures the full speech band. This is what every modern speech model is trained on.
- 44.1 kHz / 48 kHz (CD and video audio) is fine but pure overhead for transcription. The model downsamples anyway.
- 96 kHz is for music mastering, not speech.
Bit depth follows the same logic. 16-bit gives you ~96 dB of dynamic range, more than enough for any voice recording. 24-bit is a production luxury.
Should you transcribe stereo or mono recordings?
Mostly mono. With one important exception.
If you have a multi-track recording where each speaker is on their own channel (a Riverside or SquadCast podcast export, a properly configured Zoom recording, a two-line phone capture), keep the channels separate. You can transcribe each track independently and get near-perfect speaker labels just from the channel mapping. No diarization model needed.
If your "stereo" file is just the same room mic copied to two channels (which is how most consumer recordings end up), collapse it to mono. You'll halve the file size with zero accuracy cost.
For more on speaker labelling specifically, see Microsoft Teams transcription with speaker labels guide. The multi-track approach there generalizes to any platform.
What about formats like M4A, FLAC, and OGG?
All fine. None of them are "worse for transcription" in any meaningful way.
- M4A (AAC) is what iPhones, Mac QuickTime, and Zoom default to. It's a good modern lossy codec; 96 kbps and up is plenty for speech.
- FLAC is lossless and about half the size of WAV. If you're archiving recordings long-term and might transcribe them again later with a better model, FLAC is the smart choice.
- OGG / Opus is what WhatsApp voice notes, Discord, and increasingly WebRTC use. Opus was engineered for voice and holds up unusually well at low bitrates.
- WMA, RealAudio, ancient AMR: convert these to anything modern before transcribing. Some libraries don't even decode them anymore.
You can always convert with ffmpeg:
ffmpeg -i input.m4a -ar 16000 -ac 1 -c:a pcm_s16le output.wav
That's the canonical "give me 16 kHz mono 16-bit WAV" incantation, which is exactly what Whisper expects.
Does converting a lossy file to WAV improve transcription?
No. This is one of the most common misconceptions we run into.
If your source is a 96 kbps MP3, re-encoding it as WAV gives you a bigger file with the same information. The compression artifacts and the frequency cutoffs are baked in. Wrapping them in a lossless container doesn't unbake them.
Where conversion does help is when:
- Your tool only accepts certain formats and you need a different container
- You want to downsample a 48 kHz file to 16 kHz to shrink it before upload
- You're collapsing unnecessary stereo to mono
Those are container or sample-rate changes, not "quality upgrades." A bad recording stays bad.
Paste any public link or upload a file and get a clean transcript in minutes. First 3 clips every month are on us — no card required.
So what should you actually do?
If you're recording fresh: 16 kHz, 16-bit, mono WAV (or FLAC if you want it smaller). That's the cleanest possible input.
If you already have files: upload what you have. Don't pre-convert unless something downstream demands it. Most tools, including ours when you transcribe an audio file, handle the common formats directly.
If you're stuck with phone-quality 8 kHz audio: accept that names and unusual words will be less reliable, and skim the transcript with the audio for any critical passages. We covered the realistic accuracy floor in transcription accuracy: what to expect.
The file format isn't where good transcripts come from. Clear speech, a decent mic, and a quiet room are. The codec just decides how much of that survives the trip.
