MP3 to Text — Transcribe Long .mp3 Files (99 Languages)

Yes. Without an account you can transcribe the first 5 minutes of any MP3 up to 200 MB. A free account raises that to 30 minutes of audio and files up to 5 GB, still with no credit card. Paid plans start at $2/month for 200 minutes (Starter), $5/month for 1,000 minutes (Basic), $10/month for 2,500 minutes (Pro), and $20/month for 6,000 minutes (Studio). All paid tiers include all 5 export formats — TXT, DOCX, SRT, VTT, and JSON.

The one published figure is 2.7% Word Error Rate — roughly 97% word accuracy — for OpenAI's Whisper Large-v3 on the LibriSpeech test-clean benchmark, reported in the official Whisper paper. That benchmark is clean read speech, so treat it as a ceiling rather than a forecast for your file. Real recordings score lower, and how much lower depends on what is wrong with the audio: a studio podcast at 192 kbps sits near the top of the range, a multi-speaker Zoom export loses ground to crosstalk, a 64 kbps phone recording loses consonant detail, and heavy background noise costs the most. We have not benchmarked these conditions against reference transcripts ourselves, so we publish the order of magnitude and the editing effort each case implies rather than invented percentages.

Audio files transit over TLS 1.2+ encryption and are stored encrypted at rest in AWS eu-west-2. Your files are not used to train AI models. Our premium transcription and speaker-labelling providers are contractually prohibited from training on customer data. The providers behind summarization, premium translation and standard transcription do not train on API inputs under their default terms. One exception worth knowing: free translation routes through a public translation endpoint that is not covered by an enterprise data-handling agreement — use premium translation if that matters for your material. We do not sell user data. You can delete files at any time from your dashboard, and account deletion is self-serve.

Yes, automatic speaker diarization is included on every transcript. Multiple speakers are labeled Speaker 1, Speaker 2, Speaker 3, and so on. You can rename speakers in the built-in editor (e.g., "Host", "Guest", actual names) before exporting.

VexaScribe transcribes MP3 files in 99 languages with automatic language detection — English, Spanish, German, Portuguese, French, Italian, Dutch, Polish, Russian, Mandarin, Japanese, Korean, Arabic, Hindi, Vietnamese, Thai, Turkish, and many more. No need to select the language manually.

Yes, every transcript opens in a synced editor where you can correct words, merge or split segments, rename speakers, and adjust timestamps. Edits save automatically. When you're ready, export to TXT, DOCX, SRT, VTT, or JSON.

There is no duration limit — the only cap is file size. Any account uploads files up to 5 GB, and 10-hour recordings are transcribed regularly. A 1-hour MP3 at 128 kbps is roughly 58 MB, so length is rarely the constraint. Without an account, the first 5 minutes of a file up to 200 MB are transcribed.

Yes, you can run OpenAI's Whisper model locally on your own machine using `pip install openai-whisper`. CPU-only processing runs at roughly 0.5–2× real-time, so a 1-hour MP3 takes 30 minutes to 2 hours on a laptop. This is the most private option for sensitive files where you want full local control. VexaScribe's cloud is considerably faster and adds speaker diarization, a polished editor, and 5 export formats.

Not reliably. Whisper's Word Error Rate on music with sung vocals exceeds 25% — the model is trained for speech, not lyrics. For studio-quality spoken-word audio (podcasts, interviews, lectures), accuracy is excellent. For music transcription, use a dedicated lyrics-recognition service.

Median processing time is 2.9 minutes per hour of audio on Standard and 1.1 minutes per hour on Premium, measured across 5,395 production jobs in October 2026. In practice that means a 1-hour podcast takes about 2 minutes on Standard and about 1 minute on Premium. The spread matters as much as the median: 90% of Standard jobs finish inside 8.2 minutes per audio hour and 90% of Premium jobs inside 8.6, so queue depth can slow either tier on a bad run. Clips under five minutes carry one to two minutes of start-up overhead, so very short files are not proportionally quicker.

Any account: up to 5 GB per file (around 90 hours of audio at 128 kbps, though 10-hour files are the longest we see in practice). Without an account: 200 MB per file, first 5 minutes transcribed. The OpenAI Whisper API has a much smaller 25 MB hard cap by comparison — VexaScribe processes large files in a single pass without you needing to chunk them.

No. The first 5 minutes of any MP3 up to 200 MB are transcribed with no account and no card, and the TXT and SRT download without signing in. A free account raises that to 30 minutes of audio, files up to 5 GB, and the full workspace.

Yes. There is no duration cap — the only limit is a 5 GB file size, and 10-hour recordings go through regularly. A 3-hour lecture takes about 7 to 8 minutes on Standard and 2 to 3 on Premium. Long files are actually proportionally faster rather than slower: a 30-to-60-minute file runs at a median 2.2 minutes per hour of audio on Standard while a file over 5 hours runs at 1.3, because the pipeline splits long audio at 90-minute boundaries and transcribes the pieces in parallel. No splitting in ffmpeg and no chunking script required, and timestamps stay continuous across the whole file.

Below about 64 kbps you lose the high frequencies that distinguish consonants like S, F and T, so expect noticeably more errors on sibilants and proper nouns than on a clean 128 kbps file — the published 2.7% error rate for Whisper Large-v3 comes from clean read speech and is a ceiling, not a prediction for compressed audio. Anything from 64 kbps up is worth uploading. Mono is fine; bitrate matters far more than channel count. Re-encoding a low-bitrate file upward does not help, because the detail is already gone.

Yes. Once the transcript exists you can translate it into any of 99 languages with timestamps preserved, so exported subtitles still line up with the original audio. Translation runs on the transcript rather than the audio, so you can transcribe once and produce several language versions from it.

All MPEG-1 Layer III variants are supported — Constant Bitrate (CBR), Variable Bitrate (VBR), and Average Bitrate (ABR) — at bitrates from 32 kbps to 320 kbps, mono and stereo. Whisper internally resamples to 16 kHz mono per the official README, so there's no need to convert your file before uploading.

No. M4P FairPlay-protected files (Apple Music DRM) and any DRM-encrypted audio cannot be transcribed — VexaScribe cannot strip DRM. If you have legal rights to the audio, re-record from the original source or purchase a DRM-free version.

MP3 compression introduces roughly 10% relative WER increase versus uncompressed WAV in noisy conditions, per IBM Watson research. OGG Opus is more efficient — only ~2% degradation. For best accuracy on long-form material, record in WAV or FLAC and convert to MP3 for storage. For short clips, MP3 at 128 kbps or higher is indistinguishable from WAV in clean environments.

Yes. VexaScribe exports every transcript as .txt by default — plain text, no formatting, one paragraph per speaker. Choose .txt in the export dropdown after transcription completes. Also available: .docx (Word with speaker headers), .srt and .vtt (subtitle files with timestamps), and .json (word-level structured data for developers).

Not directly — VexaScribe exports .txt, .docx, .srt, .vtt, and .json. For a PDF, export to .docx and use Word, Google Docs, or Pages: File → Save As / Download as → PDF. This preserves speaker labels and timestamps and gives you full control over margins, fonts, and headers.