Bulk Transcription — 50 Files per Upload, 200 per Export, One Flat Rate

The essentials

• ●50 files per upload, 200 transcripts per export. Two different ceilings, and the difference matters: a 200-interview project is four uploads but a single export. Mixed audio and video formats throughout. No daily cap beyond your plan's monthly minutes.
• ●Two quality tiers, 99 languages either way. Standard costs one minute per audio minute and detects language per file; Premium costs two and is worth it on hard audio. We don't publish a single accuracy figure — see Standard or Premium for which to pick and languages for what actually moves accuracy.
• ●ZIP delivery with original filenames preserved. One output format per export job — run it again for a second format.
• ●Flat-rate $2–$20/month. At 100h/month: $20 vs $1,500 Rev AI vs $11,940 Rev Human. For pure-API consumers, AssemblyAI batch still wins at $15.
• ●No AI training on user data, committed contractually. Data residency and sub-processor detail is in our DPA rather than on this page.
• ●Honest scope: UI-first today. Drag-and-drop, dashboard, ZIP download. No programmatic API yet — if you need one, AssemblyAI or Deepgram are stronger fits.
• ●DPA + NDA available for research labs (IRB), legal firms (privileged data), and corporate procurement.

Who actually needs bulk transcription

“Bulk transcription” is a B2B-leaning query — almost no consumer searches it. The six personas below cover ~95% of real bulk customers. Each has a distinct pain that shapes which tool wins.

Across all six personas, the common pain is the same in different language: per-minute pricing destroys budget predictability as volume grows, one-by-one uploads waste hours, compliance documentation is needed for procurement, and consistent output matters for downstream automation. Bulk transcription is not a discount feature — it's a workflow feature.

How VexaScribe bulk works (3 steps)

1. 1

   Drag a folder

   Select up to 50 files at once from your desktop, or drop a folder onto the upload area. Larger projects just repeat this step — the 50 applies per upload, not per project. Mixed formats are fine: MP3 + M4A + MP4 + WAV in the same batch. Audio is extracted from video automatically.

2. 2

   Pick languages, formats, options

   Auto-detect language per file (recommended) or force a single language. Choose output formats — pick one or several: TXT, DOCX, SRT, VTT, PDF, JSON. Toggle diarization on/off.

3. 3

   Walk away, return to a ZIP

   Processing happens in parallel. The dashboard shows per-file progress. When it is done, select what you want and export up to 200 transcripts as a single ZIP with the original filenames preserved. Partial-batch download is supported if some files fail.

Typical processing speed: 5–10 minutes per hour of audio under normal load. A 50-file batch of one-hour interviews completes in roughly 60–90 minutes (parallelized) — not 50 hours of sequential processing.

The math: flat-rate vs per-minute (100-hour project)

The load-bearing question for buyers. We picked 100 hours because it's the realistic size of one research study, one media archive sprint, or one quarter of a marketing agency's podcast workload. Competitor figures are each vendor's published list price as of writing — worth re-checking before you budget against them, since they change.

Batch mechanics: queues, parallelism, failure handling

The boring details that matter when you're processing a few hundred files and some of them are not clean.

• ●Mixed-format support. MP3, M4A, WAV, FLAC, OGG, OPUS, MP4, MOV, MKV, WEBM, AVI in one batch. Audio is extracted from video automatically — no pre-conversion needed.
• ●Parallel processing. Multiple files transcribe simultaneously, not sequentially. A 50-file batch of one-hour interviews finishes in 60–90 minutes total, not 50 hours.
• ●Per-file status dashboard. Each file shows: queued, uploading, processing, complete, or failed-with-reason. Refresh anytime; nothing is lost if you close the tab.
• ●Independent failure handling. If 3 of 50 files fail (corrupted header, no audio track, exceeds 10-hour cap), the other 47 finish normally. Failed files retry once free; you can fix and re-upload manually after that.,Exports report what they skipped. A finished export tells you how many transcripts made it into the ZIP and lists any it left out, so you never have to diff a 200-file archive against your selection to find out something is missing. On a translated export you also get the counts: served from cache, translated fresh, and failed.,
• ●Speaker labels — honest limit. Diarization is consistent within each file (Speaker 1 through Speaker 10) but NOT auto-matched across files. The bulk-rename UI applies consistent labels across all files in a batch in one operation — typical research study with one interviewer and many participants takes ~30 seconds total to relabel.
• ●Partial-batch download. Don't wait for stragglers — download a ZIP of completed files anytime. Failed files can finish later and be downloaded separately.

Standard or Premium, and why it matters more on a batch

You pick a transcription quality when you upload, and on a large batch it is the most expensive decision on the page. Standard spends one minute of your balance per minute of audio. Premium spends two. Both cover the same 99 languages, so this is not a coverage question — it is a question of how hard your audio is.

The upload dialog shows the arithmetic before you commit: it adds up the batch and tells you what it will consume against your remaining balance, so a 200-minute batch reads as 200 minutes on Standard and 400 on Premium. Worth checking on a big run, because the difference is a doubling rather than a rounding error.

Our honest guidance: start on Standard. For clear single-speaker audio — a podcast with a decent microphone, a lecture recording, a one-on-one interview in a quiet room — the extra minutes usually buy you very little. Premium earns its cost on the audio that is genuinely hard: heavy background noise, strong accents the model sees rarely, several people talking across each other, or a recording you cannot redo. The 30 free minutes exist for exactly this test — run your worst file both ways and compare before committing a 200-file batch.

The upload dialog totals the batch against your balance before you commit, so the 1× versus 2× difference is visible rather than something you work out afterwards. Language and speaker count are set here too.

The quality tier is fixed when the file is processed. Changing your mind means re-uploading and spending the minutes again, which is the other reason to test on one file first rather than discovering it at the end of a large job.

Output formats and what lands in the ZIP

Pick a format and the whole selection comes back in it, up to 200 transcripts per export job. The ZIP preserves your original filenames with the extension matched to the output (interview.mp3 → interview.docx, interview.srt, and so on). Available formats: TXT, DOCX, SRT, VTT and JSON — one per export job, so run it again if you need the same batch in a second format.

Filenames are the thing that makes a 200-file ZIP usable: interview-maria-2026-06-10.mp3 comes back as interview-maria-2026-06-10.srt, so a script can map outputs to sources without a lookup table. If you need structured per-file data for a search index or a database, export as JSON — it carries the transcript with its timings rather than just the text.

Selection is capped at 200 per export job, and format is a single choice rather than a multi-select — which is why a batch you want in two formats means running the export twice.

Three export settings are worth knowing before you run a large job, because they apply to the entire selection at once. Speaker names lets you map Speaker 1 and Speaker 2 to real names across the whole batch rather than editing each file. Word-level timings in the JSON export give you per-word timestamps for anything you intend to process programmatically. And speaker labels and timestamps can each be turned off, which is what you want when the transcripts are headed for analysis software that expects clean prose.

Translating a whole batch

Translation happens at export, not at upload — which matters for how you plan a project. You transcribe once in the source language, then export the same set as many times as you need: the original, a translated set for a partner, another language for a client. The transcription minutes are spent once.

The flow is: select the transcripts, pick a format, then turn on translation and choose one of 133 target languages along with whether to keep speaker labels and whether to merge consecutive lines from the same speaker. Those toggles apply to every item in the batch. The cost appears on the step after that, before anything is charged. One target language per export job; run the export again for a second language.

Translation is a toggle on the export itself, and every setting here applies to all 148 items. Cost appears on the following step, before anything is charged.

You see the cost before you commit

This is the part worth knowing if you are budgeting a large multilingual project. Premium translation costs minutes, so before anything is charged a preflight runs and returns a per-file breakdown: character count and minute cost for each transcript, the total, how many files are already cached, and whether your current balance covers it. If it does not, it names the plan that would. Nothing is deducted and no job is created until you confirm.

After the export finishes you get the matching summary: how many were served from cache, how many were translated fresh, how many failed, and the total minutes actually deducted. A translation that errors server-side refunds its minutes automatically, so a partial failure does not quietly cost you.

Standard or Premium

Both engines are available on every plan. Standard is free and does not consume minutes. Premium costs minutes and reads each line more carefully — proper names and place names stay as names instead of being translated into words, idioms come out as natural phrasing rather than word-for-word, and the speaker's register is preserved. For a research corpus or anything a client will read, Premium is usually worth the minutes; for a rough gist of what is in a file, Standard is enough. Bulk export uses the Premium engine, which is why it carries a cost preflight at all.

Translating a single transcript, or want the mechanics in more depth? See the transcribe-and-translate guide. Note the honest boundary: this translates transcripts, so the audio stays in its original language — there is no dubbing or voice-over step.

Data handling, and what we'll put in writing

What procurement, legal, and IT usually ask first. The commitments below are the ones we can state publicly; anything specific to our infrastructure or sub-processors lives in the DPA rather than on a marketing page.

We deliberately do not publish our infrastructure, vendors, or processing locations on this page. Everything a security review needs — every sub-processor with its role, where each processing step happens, and the jurisdictions involved — is set out specifically in the DPA, which we will send before you upload anything. One thing worth raising early rather than late: if you work in an adversarial legal context, where the question is not just who can access data but which courts can compel it, tell us at the start. That constraint changes what we recommend, and it is a much cheaper conversation before a 200-file batch than after one.

Languages, accuracy, and what actually moves it

Transcription runs on Whisper Large-v3 Turbo across 99 languages. A bulk batch can mix languages freely — each file is detected independently from its first 30 seconds, so a folder of English, Portuguese and Japanese interviews needs no sorting before upload.

We group languages into tiers rather than publishing a percentage per tier. The grouping follows how much training data the model saw for each language, which is the documented driver of the difference. The percentages we used to print here were not measurements we made against a labelled reference set, so they are gone — and on a page aimed at people about to commit a 200-file project, an invented number is worse than none.

What reliably costs you accuracy is the recording, not the language: music under speech, two people talking over each other, a phone on a meeting table rather than near the speaker, and proper nouns or jargon the model never saw. The 30 free minutes on signup exist for exactly this — run the worst file in your set first, because that tells you more about your project than any benchmark table.

Notable PT-BR differentiator: Otter.ai does NOT support Portuguese in 2026 — its official supported languages are English, French, and Spanish only. For Brazilian agencies, Portuguese-language researchers, and LATAM media operations, VexaScribe is the practical choice over Otter for bulk Portuguese workloads. We cover this in depth in our PT-BR transcription guide.

Methodology and sources

• ● Competitor pricing taken from vendor sites as published: Rev.com, Otter.ai, AssemblyAI, Deepgram, Trint, OpenAI Whisper API. Pricing changes — always confirm on the source.
• ● Whisper Large-v3 (the base model our Turbo variant derives from): OpenAI, November 2023. MIT license. Paper: Radford et al. “Robust Speech Recognition via Large-Scale Weak Supervision” (2022).
• ● GDPR: Regulation (EU) 2016/679, Article 28 (processor obligations).
• ● Otter.ai PT-BR support: verified on otter.ai/languages June 2026 — English, French, Spanish only.