By whispertranscribe.ai Editorial Team · Last updated: November 2026 · 12 min read
The best audio format for transcription is lossless audio, WAV (16-bit PCM) or FLAC, recorded in mono at 16 kHz or higher, when you control the recording. If you already have an MP3, M4A, or video file, upload it as it is. Converting a compressed file can't bring back lost detail, and microphone distance and noise affect results more.
Below: what to choose before you press record, what to do with a file you were handed, file-size formulas, settings by scenario, FFmpeg commands, and a fix-it table.
Disclosure. whispertranscribe.ai publishes this guide. Technical claims link to Google Cloud, OpenAI's Whisper code, FFmpeg, or the Library of Congress. We ran no accuracy tests for this article, and nothing here promises that one format adds a set percentage of accuracy.
Quick answer: the best audio format for transcription
Pick by where you are in the workflow, not by a single "winner."
| Your situation | Best choice | Why |
|---|---|---|
| You're setting up a recorder or app | WAV (16-bit) or FLAC, mono, 16 kHz or higher | Google recommends lossless FLAC or LINEAR16 for speech recognition |
| You need small files while recording | A speech-friendly lossy codec at a moderate bitrate | Google's preferred lossy codecs for speech, in order: AMR-WB, Ogg Opus, Speex |
| You already have an MP3, M4A, or Opus file | Upload the original | Re-encoding adds nothing back |
| You have a video (MP4, MOV, MKV) | Upload it; extract audio if it's huge | Only the audio track is used |
| Speakers were recorded on separate tracks | Keep the tracks | Google advises sending each channel separately when each person has their own |
When not to worry about format at all: a phone voice memo, a published podcast episode, or a Zoom or Teams recording. Upload it, pick the right spoken language, and spend your time checking names instead.
Format, codec, and container: what the file extension hides
The extension names the container, not always the compression. Google's audio optimization tutorial says it plainly: codecs and file formats are not the same. Three common surprises:
- WAV usually holds uncompressed linear PCM, but the format can carry other encodings too (Library of Congress, WAVE).
- M4A can hold lossy AAC or lossless Apple Lossless (ALAC).
- OGG is a wrapper that usually holds Opus or Vorbis, both lossy.
To see what's actually inside a file, ask FFmpeg's ffprobe for the audio streams only:
ffprobe -v error -select_streams a -show_entries stream=index,codec_name,channels,sample_rate -of csv=p=0 meeting.mp4
We ran this with FFmpeg 7.1.5 on a short test video: it printed one line per audio stream. On a copy with the audio removed, it printed nothing, a quick way to catch a file with no audio track before uploading.
Audio formats compared for transcription
Every format below can be transcribed. What differs is how much detail survived and how big the file gets. Sizes come from the formulas in the next section (1 MB = 1,000,000 bytes).
| Format | Lossy or lossless | Where it usually comes from | Size per minute (example setting) | Effect on transcription | What to do |
|---|---|---|---|---|---|
| WAV (PCM) | Uncompressed | Field recorders, audio editors | 1.92 MB at 16 kHz, 16-bit, mono; 10.58 MB at 44.1 kHz, 16-bit, stereo | Keeps everything the mic captured | Best master; watch size on long sessions |
| FLAC | Lossless (compressed) | Archives, editor exports | Smaller than the same WAV; saving varies | Same audio as its WAV | Use instead of WAV to save space |
| AIFF / ALAC / CAF | AIFF usually uncompressed; ALAC lossless; CAF can hold either kind | Mac and iPhone apps | AIFF: same as WAV at the same settings | Depends on the codec inside | Upload as is |
| MP3 | Lossy | Podcasts, downloads | 0.96 MB at 128 kbps | Google: lossy codecs may reduce accuracy | Upload as is |
| M4A / AAC | Usually lossy (AAC) | Voice memos, video soundtracks | 0.48 MB at 64 kbps | Same caution as MP3 | Upload the original |
| OGG / Opus | Lossy | Messaging voice notes, web calls | 0.24 MB at 32 kbps | Opus is on Google's preferred list for lossy speech | Upload as is |
| AMR (narrowband) | Lossy, speech-only | Older phones and call recorders | About 0.09 MB at 12.2 kbps | Carries only 200 to 3,400 Hz (Library of Congress, AMR) | Upload as is; never upsample |
Video files work the same way: the tool reads the audio track, so the only cost is extra upload size.
How big is an audio file? The formulas
Two formulas let you estimate any file size before you record or upload.
Uncompressed PCM (WAV, AIFF):
bytes per second = sample rate × (bit depth ÷ 8) × channels
Worked examples:
- 16 kHz, 16-bit, mono: 16,000 × 2 × 1 = 32,000 bytes per second. × 60 = 1,920,000 bytes, about 1.92 MB per minute, or 115.2 MB per hour.
- 44.1 kHz, 16-bit, stereo: 44,100 × 2 × 2 = 176,400 bytes per second, about 10.58 MB per minute, or 635 MB per hour.
- 48 kHz, 24-bit, stereo: 48,000 × 3 × 2 = 288,000 bytes per second, or 1,036.8 MB per hour.
Lossy at a constant bitrate (MP3, AAC, Opus):
bytes per minute = bitrate in kbps × 1,000 ÷ 8 × 60 (that is, kbps × 7,500)
- 128 kbps: 128 × 7,500 = 960,000 bytes, about 0.96 MB per minute, or 57.6 MB per hour.
- 64 kbps: 0.48 MB per minute; 32 kbps: 0.24 MB per minute.
Real files run slightly larger because of headers, and variable-bitrate files drift around the average. Computers that show binary units (MiB) display about 5% lower numbers.
What those numbers mean for a 5GB, 10-hour upload limit
whispertranscribe.ai accepts files up to 5GB and 10 hours. Here's how common settings fit, assuming 5GB means 5,000,000,000 bytes:
| Setting | 1 hour | 10 hours | Fits 10 hours under 5GB? |
|---|---|---|---|
| WAV, 16 kHz, 16-bit, mono | 115.2 MB | 1.15 GB | Yes |
| WAV, 44.1 kHz, 16-bit, stereo | 635 MB | 6.35 GB | No: about 7.9 hours fit |
| WAV, 48 kHz, 24-bit, stereo | 1.04 GB | 10.37 GB | No: about 4.8 hours fit |
| MP3, 128 kbps | 57.6 MB | 576 MB | Yes |
| Opus, 32 kbps | 14.4 MB | 144 MB | Yes |
A classic WAV file also has its own ceiling: its 32-bit size field limits it to about 4 GB, and longer recordings need the RF64 extension (Library of Congress). For very long lossless recordings, FLAC or a 16 kHz mono WAV avoids both problems.
What matters more than the audio format
A clear voice close to the microphone, saved as MP3, usually beats a lossless file recorded across a noisy room. These settings and habits decide most results.
Sample rate: is 16 kHz enough?
Yes, for speech. Google recommends 16,000 Hz or higher (best practices); its tutorial explains that lower rates leave little or no information above 8 kHz. OpenAI's open-source Whisper code converts every input to 16,000 Hz mono before the model hears it (whisper/audio.py). Recording at 44.1 or 48 kHz is fine, just bigger.
Raising the rate afterwards doesn't help: Google notes that an 8 kHz file upsampled to 44.1 kHz still holds only about 4 kHz of frequency range. Upload telephone audio, commonly 8,000 Hz, at its original rate.
Bit depth: 16-bit or 24-bit?
Google recommends 16 bits or more and says converting 8-bit audio to 16-bit gains nothing, because the dynamic range was fixed at recording. Keeping 24-bit from your recorder is harmless; converting 16-bit to 24-bit adds nothing.
Bitrate for compressed files
No standard sets a bitrate that guarantees good transcription. Google's guidance is about codec type: lossy codecs may reduce accuracy, and noise or echo hurt more when lossy compression is involved. If you must compress, do it once at a moderate setting and avoid chains like WAV to MP3 to OGG.
Microphone distance, noise, and levels
From Google's preprocessing advice:
- Put the microphone as close to the speaker as you reasonably can, especially with background noise.
- Avoid clipping. Distorted peaks can't be repaired later.
- Turn off automatic gain control and noise reduction if your app allows. Google says noise-reduction processing before recognition typically reduces accuracy.
Mono or stereo?
For one shared microphone, mono is enough. Whisper mixes stereo down to mono anyway, so stereo with the same voice on both sides only doubles the upload. Stereo is worth keeping when each side holds a different person, as with many call recorders or one lavalier per channel. Google advises sending each channel separately then, and a mixed recording as it is.
Multitrack recordings
Podcast and remote-interview tools often record each person on a separate track. Either export one mixed track and turn on speaker labels (simplest), or transcribe each track separately and merge the texts by timestamp (cleanest audio, but you do the merging).
Recommended format and settings by scenario
"If you already have" is the more common situation, and the answer is usually "upload it."
| Scenario | Record as (if you choose) | If you already have | Before uploading | Watch out for |
|---|---|---|---|---|
| In-person interview | WAV or FLAC, mono per mic, 44.1 or 48 kHz | M4A or MP3 from a phone | Note names and terms | Guest quieter than interviewer |
| Remote interview or podcast, per track | Each speaker's own track, lossless | Separate WAVs plus a mixed MP3 | Pick mixed or per-track | Crosstalk on the mix |
| Meeting recording (Zoom, Teams, Meet) | The platform's audio option, if offered | MP4 or M4A | Extract audio only if the video is huge | One laptop mic for a whole room |
| Phone voice memo | The app's default | .m4a | Export the original | Phone in a pocket while recording |
| Published podcast episode | Not applicable | The released MP3 | Use the pre-release WAV or FLAC if you have it | Music intros; set the language |
| Lecture or seminar | Mono, 16-bit, mic near the lecturer | A long MP3 or video | Check length against 10 hours | Audience questions off-mic |
| Phone call recording | The recorder's default | AMR, M4A, or 8 kHz WAV | Keep the original sample rate | Consent rules; don't upsample |
Recording calls raises consent questions; our guides on how to record a phone call and call recording laws by state cover them. This is not legal advice.
How to convert audio for transcription with FFmpeg
Convert only for a reason: the file is too big, you want a lighter upload from a video, a tool rejects the codec, or you want speakers in separate files. Converting for "quality" doesn't work.
We ran each command with FFmpeg 7.1.5 on an 11-second public-domain speech clip and a short test video made from it, checking only that it finished and that ffprobe showed the expected codec, sample rate, and channels. Options are explained in the FFmpeg documentation.
1. Any file to 16 kHz mono 16-bit WAV (the same shape Whisper converts to internally):
ffmpeg -i interview.m4a -ac 1 -ar 16000 -c:a pcm_s16le interview_16k.wav
2. Any file to 16 kHz mono FLAC (lossless and smaller than the WAV in most cases):
ffmpeg -i interview.m4a -ac 1 -ar 16000 -c:a flac interview_16k.flac
3. Video to a small mono MP3 (adds one lossy step; use it when size matters most):
ffmpeg -i meeting.mp4 -vn -ac 1 -ar 16000 -c:a libmp3lame -b:a 64k meeting_16k.mp3
4. Pull the audio out of a video without re-encoding (works when the audio is AAC, which is common in MP4 files):
ffmpeg -i meeting.mp4 -vn -c:a copy meeting_audio.m4a
5. Split a two-channel recording into one file per speaker (uses the channelsplit filter):
ffmpeg -i call.m4a -filter_complex "channelsplit=channel_layout=stereo[L][R]" -map "[L]" left.wav -map "[R]" right.wav
| Option | What it does |
|---|---|
-ac 1 |
Mixes the audio down to one channel |
-ar 16000 |
Sets the output sample rate to 16,000 Hz |
-c:a pcm_s16le |
Writes 16-bit uncompressed PCM (standard WAV) |
-vn |
Drops the video stream |
-c:a copy |
Copies the audio as is, with no quality change |
-b:a 64k |
Sets the lossy bitrate |
Prefer command 4 over command 3 when the video's audio is already AAC, since it keeps the original quality. For 8 kHz phone audio, leave out -ar 16000 so you don't upsample.
No command line? whispertranscribe.ai's audio and video to MP3 converter extracts the audio track from MP4, MOV, MKV, WEBM, or AVI in your browser. The page says it uses WebAssembly and that files are processed in the browser, not uploaded.
Large files: staying under 5GB and 10 hours
- Too big but under 10 hours: usually a video or a high-resolution WAV. Extract the audio (command 4 or the browser converter) or convert the WAV to FLAC (command 2).
- Longer than 10 hours: the upload page shows "At the moment, we can only process files up to 10 hours long. Please shorten your file to 10 hours or less and try again." Split it. The no-re-encode split command in our MP3 to text guide also works on WAV; we checked it on a test WAV. Check the joins when you merge text.
- Slow connection: one hour of 64 kbps MP3 is about 28.8 MB, against 635 MB for CD-style stereo WAV. Keep the lossless original as your archive.
Troubleshooting audio format problems
| Problem | Likely cause | What to try |
|---|---|---|
| "No audio detected. The file may be corrupted. Please check the file and try again." (shown on our upload pages) | Damaged file, a video exported without sound, or an empty audio track | Play it locally; run the ffprobe audio check above (no output means no audio stream); re-export from the original app |
| Upload refused as too long | File over 10 hours | Split into parts of 10 hours or less |
| File too large | Video or high-resolution, multichannel WAV | Extract the audio or convert WAV to FLAC |
| A WAV over 4 GB won't open or upload | Classic WAV size limit | Re-export as RF64 or FLAC, or split it |
| One speaker is much quieter than the other | Each person is on a separate channel, recorded at different levels | Split the channels (command 5) and transcribe each file |
| Distorted, crackly peaks | Clipping during recording | Can't be repaired; lower input gain next time |
| Phone audio still poor after converting to 48 kHz | Upsampling adds no information | Upload the original; try a "Noise / Low-Quality Audio/Video" language option where offered |
Where whispertranscribe.ai fits
whispertranscribe.ai is a transcription tool powered by OpenAI's Whisper models that runs in a web browser (no app, no API, no human proofreading). The free audio to text converter and other upload pages list these formats: MP3, WAV, M4A, AAC, FLAC, OGG, OPUS, AMR, AIFF, ALAC, CAF, MP4, MOV, M4V, MKV, AVI, WMV, WEBM, MPG, and MPEG.
- Limits: files up to 5GB and 10 hours.
- Languages: 130+ languages, with "Noise / Low-Quality Audio/Video" options for many of them.
- Accuracy: up to 99% on clear audio. Noise, distance, accents, and crosstalk lower it.
- Speakers: turn on Speaker Detection before transcribing.
- Exports: DOCX, PDF, TXT, CSV, SRT, and VTT.
- Privacy: the privacy policy says original uploads are deleted after processing and extracted MP3 files are kept for 30 days by default.
For specific recordings, see the pages for MP3 files, voice memos, phone calls, interviews, and lectures and seminars. Try a file on the no-account free page (limits are shown there). Paid plans: from $9.99/month billed annually (pricing). Everything else is on the features page.
FAQs
What is the best audio format for transcription?
WAV or FLAC is the best audio format for transcription when you control the recording, ideally in mono at 16 kHz or higher with 16-bit depth. When you already have a compressed file such as MP3, M4A, or Opus, upload it as it is rather than converting. Recording quality, meaning a close microphone, low background noise, and no clipping, usually matters more than which of these formats you pick.
Is WAV better than MP3 for transcription?
A WAV recorded from the source keeps more detail than an MP3, and Google notes that lossy codecs like MP3 may reduce recognition accuracy. The gap is often small on clear speech and larger on noisy recordings. A WAV made from an existing MP3 is no better than that MP3, because the discarded detail is gone. Record in WAV or FLAC when you can, and upload MP3s as they are.
What sample rate should I use for transcription?
Use 16 kHz or higher. Google recommends at least 16,000 Hz for speech recognition, and OpenAI's Whisper converts all audio to 16 kHz mono before processing. Recording at 44.1 or 48 kHz is fine and common. Never raise a lower rate afterwards: an 8 kHz phone recording should be uploaded at 8 kHz, because upsampling adds file size but no speech information.
Should I record in mono or stereo for transcription?
Record in mono when everyone shares one microphone. Whisper mixes stereo to mono anyway, so a stereo copy of the same voice only doubles the file size. Keep stereo or multitrack when each speaker is on a separate channel, such as a call recorder or one microphone per guest. Google advises transcribing separate channels separately, which you can do by splitting the channels with FFmpeg.
How big is a WAV file per minute?
It depends on the settings: multiply sample rate by bytes per sample by channels to get bytes per second. A 16 kHz, 16-bit mono WAV is 16,000 × 2 × 1 = 32,000 bytes per second, about 1.92 MB per minute or 115.2 MB per hour. CD-style 44.1 kHz, 16-bit stereo is about 10.58 MB per minute, or 635 MB per hour.
Can I upload a video file instead of extracting the audio?
Yes. Transcription tools read the audio track and ignore the picture, so an MP4, MOV, MKV, or WEBM gives the same transcript as its extracted audio. Video just makes the upload larger. Extract the audio when a file is close to a size limit or your connection is slow, for example with FFmpeg's -vn -c:a copy, which keeps the original audio quality. On whispertranscribe.ai you can upload the video as it is and export DOCX, TXT, SRT, or VTT.
Have a recording ready? Upload it as it is to whispertranscribe.ai on the free audio to text converter: pick one of 130+ languages, switch on Speaker Detection, then download DOCX, TXT, SRT, or VTT. Files up to 5GB and 10 hours.
References
- Google Cloud: Speech-to-Text best practices (providing speech data)
- Google Cloud: Optimize audio files for Speech-to-Text
- OpenAI: whisper/audio.py in the Whisper GitHub repository
- FFmpeg: ffmpeg command-line documentation and channelsplit filter
- Library of Congress: WAVE Audio File Format (fdd000001)
- Library of Congress: AMR, Adaptive Multi-Rate Speech Codec (fdd000254)
