How accurate is AI audio transcription?

On clean studio audio with one English speaker, modern services hit 95–98% word accuracy (2–5% WER). On a Zoom call with 4 speakers and mixed mics, expect 80–90%. On phone call audio (narrow-band codec), 70–85%. On audio with music or heavy background noise, often much worse. The '99% accurate' marketing claim is real only on the easiest condition — see the WER-by-condition table below.

Is there a truly free option?

Yes. iPhone Voice Memos on iOS 18+ transcribes on-device for free (English, plus a growing list of languages). Pixel Recorder is free on Pixel phones. Microsoft 365 subscribers get Word's Transcribe feature included. Self-hosted Whisper is free forever (requires Python and decent hardware). Cloud services offer free tiers — DeluxeScribe gives 60 minutes free with no credit card. "Free web tools" typically cap at 5–10 minutes before paywalling.

Can I transcribe an audio file over 3 hours long?

Yes. DeluxeScribe accepts files up to 5 GB — roughly 90 hours of audio at typical bitrates. Most cloud services handle multi-hour files without splitting. Self-hosted Whisper has no length limit (limited only by your disk and time). The bottleneck is usually upload speed on cloud services, not processing.

What file formats work?

Direct upload: MP3, M4A, WAV, MP4, MOV, FLAC, AAC, OGG, Opus, WebM, and 15+ others. May need conversion first: MKV (re-mux to MP4), AMR (convert to M4A), some proprietary formats. The original file format barely affects accuracy — audio quality at recording time matters more than the container.

Can I get speaker labels?

Yes, on cloud services and with Whisper + a diarization add-on (WhisperX or whisper-diarization). DeluxeScribe includes automatic speaker labels. Expect 85–95% attribution accuracy on clean recordings, dropping to 70–85% on Zoom-style mixed-mic setups. iOS 18's Voice Memos transcription does NOT include speaker labels.

How long does transcription take?

Cloud services typically finish a 1-hour file in 3–10 minutes. Self-hosted Whisper on a CPU takes 10–30× real-time (a 1-hour file takes 10–30 hours). On a recent GPU, near real-time. Native OS transcription (iPhone, Pixel) processes in the background at roughly real-time speed.

Does audio quality matter more than bitrate?

Yes, dramatically. A 64 kbps recording from a close microphone in a quiet room outperforms a 320 kbps recording from a phone across a noisy room. Mic distance, room acoustics, and background noise dominate accuracy. Bitrate matters only below ~32 kbps where compression artifacts start hurting speech recognition.

Is my audio kept private?

Depends on the path. Cloud services: audio is uploaded, processed, and stored per the vendor's retention policy. DeluxeScribe encrypts in transit and at rest, doesn't train on customer audio, and lets you delete anytime. Self-hosted Whisper: nothing leaves your machine. Native OS on iPhone/Pixel: processes on-device. For medical, legal, or NDA-covered content, self-host — no cloud service (including ours) is HIPAA-compliant.

Can I transcribe non-English audio?

Yes. DeluxeScribe supports 99 languages. Whisper (self-hosted) supports 99+. Most cloud services support 30–60 languages. Native OS options are more limited: iOS 18 launched with English and added more in point releases; Word Transcribe supports ~10 languages; Pixel Recorder supports a handful.

Transcribe Audio: 4 Real Paths (2026)

Upload to a cloud service, use a free tier, run Whisper locally, or use your OS built-in — the honest decision framework with real accuracy figures.

Four paths exist:upload to a cloud service (easiest, paid above a free tier), self-host Whisper (free, technical), use a free tier (has limits), or use your operating system’s built-in transcription (macOS Voice Memos, Microsoft Word, Pixel Recorder). Pick based on the file you have, the accuracy you need, and whether you can upload the audio. DeluxeScribeis one of the cloud options — 99 languages, files up to 5 GB, speaker labels, six export formats. Free tier is 60 minutes; paid plans start at $10/month. Below: the decision framework, honest accuracy by audio condition (not the usual “99%” blanket claim), pricing across the field, and format-specific routing for your exact case.

60 minutes free
No credit card
99 languages
Speaker labels

Last verified July 8, 2026

TL;DR — pick your path

Match your situation to a path. Each is explained in detail below.

Your situation	Best path
Any audio file, want it done fast	Cloud upload — DeluxeScribe (free 60 min)
Sensitive content — can’t upload to cloud	Self-hosted Whisper (local, free, same accuracy)
Long recording (over 3 hours) or ongoing volume	Paid cloud tier or self-hosted Whisper on GPU
iPhone (iOS 18+) with a short English clip	Voice Memos built-in transcription (on-device)
You have a Microsoft 365 subscription	Word Transcribe (free with subscription)
Format-specific need (MP3, WAV, M4A, MP4)	Jump to the format page (linked below)
Building transcription into your product	AssemblyAI, Deepgram, or self-hosted Whisper API

How to transcribe audio (3 steps)

Upload — drop your file into DeluxeScribe or any cloud transcription service. MP3, M4A, WAV, MP4, and 20+ other formats work directly.
Select language — pick the spoken language or leave on auto-detect (runs on the first ~30 seconds).
Export — download as TXT, DOCX, PDF, SRT, VTT, or JSON with speaker labels and timestamps.

That’s the common case. If your situation is different (sensitive content, ongoing volume, iOS built-in), the sections below cover it.

Transcribe any audio file in 99 languages

60 minutes free, no credit card. Speaker labels, six export formats, files up to 5 GB.

The 4 paths, compared

Path	Cost	Speed	Privacy	Max length	Languages	Speaker labels
Cloud service (DeluxeScribe et al.)	Free tier → $10+/mo	5–10 min per hour	Cloud-encrypted	5 GB / ~90 hrs	99	Automatic
Self-hosted Whisper	Free	CPU: 10–30× realtime, GPU: near realtime	Local (never uploads)	Your disk	99	Add-on (WhisperX)
Free web tools	Free (with caps)	5–15 min per hour	Cloud (varies)	Usually 30 min	10–40	Sometimes
Native OS (iOS/macOS/Word/Pixel)	Free / included	Realtime (background)	On-device (mostly)	Varies (iOS: ~30 min)	1–12	No

Path 1 — Cloud transcription service

Upload audio in a browser, get a transcript back in minutes. The easiest path for most use cases. Free tiers exist; paid plans start around $10/month.

When cloud wins: multiple files, non-English audio, need speaker labels without setup, need in-browser editing to fix errors, want SRT/VTT/DOCX/JSON exports, need speed.

When cloud loses: sensitive content (medical, legal, NDA), offline requirement, cost sensitivity at high volume.

Path 2 — Self-hosted Whisper

OpenAI’s Whisper model runs locally with one command if you have Python. Audio never leaves your machine.

pip install openai-whisper
whisper recording.mp3 --model large-v3 --output_format srt

Model size tradeoff: tiny and base are fast but rough; medium is a good default; large-v3 is state-of-the-art but ~10× slower on CPU.

Speed reality:on a typical CPU, expect 10–30× real-time (a 1-hour file takes 10–30 hours). On a recent GPU it’s real-time or faster. For speaker labels, pair with WhisperX (adds word-level timestamps + diarization) or faster-whisper for a 4× speedup.

When Whisper wins: medical, legal, or confidential audio; high volume where cloud cost adds up; you already have a GPU.

Path 3 — Free web tools (with caveats)

Sites like audiototext.com, freepodcasttranscription.com, audioconvert.ai offer browser-based transcription with no signup. Convenient for a one-off short file.

The catch:“free” almost always hides a paywall. Common patterns: file-length cap (10–30 minutes), transcript truncation after the free portion, watermarked output, paywall appearing after processing.

Honest DeluxeScribe note:our free tier gives you 60 minutes upfront with no card required — no after-the-fact paywall, no watermark, all export formats included. It’s the fairest free tier we’ve found; we’re not saying that’s neutral, we’re saying that’s why the “free web tools” category is often worse than a legitimate free tier.

Path 4 — Native OS options

Your operating system probably has built-in transcription. All free, all language-limited relative to cloud services.

iPhone Voice Memos (iOS 18+) — on-device transcription, English + growing language list, iPhone 12+. See our iPhone Voice Memo guide.
macOS Voice Memos — same engine as iOS, syncs via iCloud, Apple Silicon Macs.
Microsoft Word Transcribe — included with Microsoft 365, web-only, supports MP3/WAV/M4A/MP4, ~10 languages. Records live or accepts file upload.
Google Docs voice typing — live only (no file upload), 100+ languages, free with Google account.
Pixel Recorder (Android) — on-device on Pixel 3+, English + a handful of other languages.

Accuracy reality — WER by audio condition

Every vendor claims “99% accuracy.” That number is real — on the easiest possible condition. On everything else, accuracy drops in predictable ways.

Word Error Rate (WER) is the standard metric: the percentage of words wrong (substitutions + insertions + deletions). 5% WER means 5 out of 100 words are wrong. Under 15% is workable with light editing.

Audio condition	Typical WER	Reads like
Studio mic, one speaker, clear English	2–5%	Almost perfect; light proofreading
Conference room, 2–3 speakers, good mics	5–10%	Useful as-is; spot-check names
Zoom call, mixed mics, 4+ speakers	10–20%	Readable, needs editing for publication
Phone call (8 kHz narrow-band)	15–30%	Gist clear, details unreliable
Heavy accent + background noise	20–40%	Re-listen to verify key points
Music + speech overlap	Often fails	Model may hallucinate lyrics or skip sections

For the full breakdown by language and Whisper model size, see How Accurate Is Whisper.

Where all providers fail the same way

Proper nouns — names, businesses, product names
Phone numbers — fast or non-standard groupings
Technical jargon — drug names, legal terms, science
Homophones — their/there/they’re, principal/principle
Numbers with units — “15 mg” vs “50 mg”

Spot-check these regardless of vendor.

Pricing reality across the field

What you actually pay per hour of transcribed audio, across the main cloud services. Verified July 8, 2026.

Service	Free tier	Paid from	Approx. $/hr audio	Notes
DeluxeScribe	60 min one-time	$10/mo · 1,200 min	~$0.50/hr	99 languages, speaker labels, 6 export formats
TurboScribe Unlimited	3 files/day	$10/mo	Effectively $0 at volume	Unlimited hours; single-tier pricing
Otter	300 min/mo	$17/mo · 1,200 min	~$0.85/hr	Meeting bot, calendar integration
Descript	1 hour/mo	$24/mo · 30 hrs	~$0.80/hr	Edit audio by editing transcript text
Rev AI (self-serve)	None	$0.25/min	$15/hr	Pay-per-minute; optional human tier at $1.50/min
Trint	None	$48/mo · 7 hrs	~$6.85/hr	Newsroom editor for journalists
Self-hosted Whisper	Free	—	$0 (electricity)	Free forever; needs hardware + Python

For a full comparison of TurboScribe alternatives, see our TurboScribe alternatives rankings. For Whisper-specific alternatives (developer/API focus), see Whisper alternatives.

File formats — what works and what needs converting

Most modern services accept every common audio and video format. The exceptions are rare and easy to convert.

Upload directly (works everywhere)

Format	Common source	Format-specific guide
MP3	Podcasts, music apps, universal	MP3 to Text
M4A	iPhone Voice Memos, WhatsApp voice notes	M4A to Text
WAV	Uncompressed studio, pro audio apps	WAV to Text
MP4	Zoom exports, screen recordings, phone video	MP4 to Text
MOV, WebM, FLAC, AAC, OGG, Opus	QuickTime, browser recordings, audiophile	Upload directly to DeluxeScribe

Convert first

Source	Convert to	FFmpeg command
MKV video	MP4 (re-mux, no re-encode)	`ffmpeg -i input.mkv -c copy output.mp4`
AMR (old Android recorder)	M4A	`ffmpeg -i input.amr -c:a aac output.m4a`
WMA (old Windows)	MP3	`ffmpeg -i input.wma output.mp3`
ALAC (Apple Lossless in .m4a)	AAC in M4A	`ffmpeg -i input.m4a -c:a aac output.m4a`

By source — jump to your exact use case

This page is the overview. For specific formats, sources, or workflows, we have dedicated guides with exact steps and gotchas.

By format

MP3 to Text — the universal format
M4A to Text — iPhone Voice Memos, WhatsApp
WAV to Text — uncompressed studio
MP4 to Text — video files

By recording source

iPhone Voice Memo Transcription — iOS 18 deep dive
Zoom Transcription — native + third-party
Interview Transcription — journalists, researchers, HR

Podcasts + long-form

Podcast Transcription — listener + podcaster workflows

Video, subtitles & captions

Video to Text — source-agnostic (upload, YouTube URL, FFmpeg)
Video to SRT — subtitle extraction + NLE workflow
Audio to SRT — audio-first subtitle workflow
SRT Generator — subtitle timing + format
Text to SRT — free frontend tool: paste text, get valid .srt (no audio needed)
VTT to SRT Converter — free browser converter for .vtt → .srt
SRT to VTT Converter — free browser converter for .srt → .vtt (HTML5 & Podcasting 2.0)

Alternatives + comparisons

Whisper Alternatives — 8 ranked options
TurboScribe Alternatives — 7 ranked by switch reason
Otter.ai Alternatives — 10 ranked by switch reason (meeting-assistant category)
Best AI Note-Taker — 10 note-takers compared by job, no #1 crown
How Accurate Is Whisper — WER by condition + language

Privacy — one honest paragraph

Cloud services:your audio is uploaded and processed on the vendor’s servers. DeluxeScribe encrypts in transit (TLS) and at rest, doesn’t use your audio to train models, and lets you delete recordings anytime. Self-hosted Whisper: nothing leaves your machine. Native OS (iPhone Voice Memos, Pixel Recorder): processes on-device.

We are not HIPAA-compliant — do not upload Protected Health Information. For PHI, use a vendor with a signed BAA, or self-host Whisper. For attorney-client privileged content, check with your firm before uploading to any cloud service.

How this page was verified

Word Error Rate (WER) ranges come from published benchmarks on the LibriSpeech corpus (clean studio audio) and the AMI Meeting Corpus (multi-speaker meeting audio), generalized across modern transformer-based ASR models. Whisper accuracy ranges align with Radford et al. (2022). Whisper command-line behavior from the OpenAI Whisper repository. Pricing figures verified against vendor pricing pages on July 8, 2026. We don’t cite “99% accurate” marketing claims that appear across competitor copy — they’re not sourced to a published study.