Bulk Transcription — 50 Files, One Flat Rate, London Storage

Up to 50 files per batch on every paid plan. The 50-file limit is generous for most workflows: a research lab with 30 hour-long interviews fits in one batch; a podcast agency with weekly client deliverables runs one batch per client. If you need more, run consecutive batches — there's no per-day or per-account cap beyond your plan's monthly minutes. The 50-file ceiling exists because larger batches degrade UI responsiveness; processing 200+ files via API is on the roadmap for developer use cases.

Per file: 5 GB and 10 hours (whichever comes first). Per batch: 50 files. There is no hard total-batch-size limit beyond the per-file cap × 50 — so a theoretical maximum batch is 50 files × 5 GB = 250 GB, though we recommend keeping batches under ~25 GB in practice for upload reliability over typical office internet. Long files (3-10 hours) work fine — common with full-day depositions, half-day workshops, or oral-history projects. For files over 10 hours, split before upload using a free tool like ffmpeg or Audacity.

Yes — mix any combination of MP3, M4A, WAV, FLAC, OGG, OPUS (audio) plus MP4, MOV, MKV, WEBM, AVI (video) in a single batch. Audio is extracted from video automatically — no need to pre-convert. Each file is detected, processed, and transcribed independently; the batch waits for all files to finish before delivering the ZIP. Mixed-format support matters for real workflows: an agency receives MP3s from one client, M4As from iPhones, MP4s from Zoom recordings — they shouldn't have to pre-process everything just to transcribe.

No. Each file is processed independently. If 3 of 50 files fail (corrupted audio, unsupported codec, exceeds duration cap), the other 47 finish and you can download a partial ZIP containing the successful files. Failed files appear in the dashboard with the specific error reason — corrupted header, no audio track detected, file exceeds 10 hours, etc. You can retry failed files individually (free retry within 24 hours) or fix the source and re-upload. The batch never silently swallows failures, and successful work is never blocked by a single bad file.

Yes. The ZIP preserves your original filenames — “interview-maria-2026-06-10.mp3” becomes “interview-maria-2026-06-10.docx” (or .srt, .txt, etc., depending on the format you selected). The ZIP also includes a CSV manifest with per-file metadata: original filename, duration, detected language, number of speakers, processing timestamp, word count. The manifest makes downstream automation easy — pipe it into a script that uploads to your CAQDAS tool (NVivo, ATLAS.ti), CMS, or shared drive without manual mapping. Multiple output formats per file are supported in the same batch (one ZIP with both .docx and .srt for every file).

Speaker labels are consistent WITHIN each file (Speaker 1, Speaker 2... up to 10) but NOT auto-matched ACROSS files. This is an honest technical limitation of diarization in 2026: cross-file speaker identification requires voice-print enrollment, which adds complexity and privacy implications we've chosen not to ship. Workaround: use the bulk-rename UI in the dashboard to apply consistent labels (Speaker 1 → “Interviewer”, Speaker 2 → “Participant”) across all files in a batch in one operation. For research studies where the same interviewer appears in all 50 files, this takes ~30 seconds total. We're transparent about this because over-promising cross-file matching is a common industry trap.

No. We contractually commit to never training models on user audio or transcripts — verifiable in our Terms and DPA. We use OpenAI's Whisper Large-v3 (open-source, MIT license) for transcription and pyannote.audio for diarization. Inference runs on specialized ML compute infrastructure (our inference partner is disclosed by name in the DPA). Our inference partner is contractually committed to not training models on inference data per their published Terms of Service. This is a deliberate differentiator vs. providers like Otter.ai, which trains on user audio by default with manual opt-out.

Storage: AWS eu-west-2 (London, UK). All audio, transcripts, and account data live in London under UK-GDPR (the post-Brexit equivalent to EU-GDPR), with AES-256 encryption at rest and TLS 1.2+ in transit. Processing: during transcription, audio is sent from our London infrastructure to an ML inference partner that runs Whisper Large-v3 on specialized GPU infrastructure, then results return to London for storage and delivery. Honest disclosure: our inference partner is US-domiciled, so the transcription step is within reach of the US CLOUD Act even though all storage is in London. For workloads where end-to-end EU residency is non-negotiable — adversarial legal contexts, sensitive journalistic sources under US legal pressure — get in touch about our dedicated infrastructure option. We retain audio for 30 days post-transcription by default; auto-delete on transcript download is available on Pro and Studio plans, and on-request immediate deletion is always supported.

Not yet. As of June 2026, bulk transcription is UI-first — drag-and-drop folder upload, dashboard status, ZIP download. Programmatic API access (POST a list of S3 URLs, receive webhooks per file completion) is on the roadmap but not shipped. If your workflow strictly requires API/webhook integration with S3 sync, we honestly recommend AssemblyAI ($0.15/hr batch) or Deepgram ($0.26/hr) — both have mature batch APIs. For everyone else (agencies, researchers, legal teams, podcasters), the UI workflow is faster than wrangling API code: drag a folder, walk away, return to a ZIP. We'll announce API access here when it ships.

For 100 hours in a single month: Rev AI charges $0.25/min = $1,500. Rev Human charges $1.99/min = $11,940. Otter Business charges $30/user/month but caps imported-file minutes at 6,000/user — 100 hours fits, so $30, but go over and you need a second seat. AssemblyAI batch is $0.0025/min = $15 (genuinely cheaper if you only need API). Trint BulkScribe starts around $0.20/hour but requires enterprise contract negotiation. VexaScribe Studio is $20/month flat for 6,000 minutes = 100 hours. The pattern: flat-rate dominates per-minute pricing once you exceed ~5 hours/month, and dominates seat-based pricing once you exceed one user's needs. For pure-API consumers, AssemblyAI may still win — we don't compete on that segment.

Yes — all 99 Whisper Large-v3 languages are supported in bulk. Tier 1 languages (92-95% accuracy on clean audio): English, Spanish, French, German, Italian, Portuguese (PT and BR), Dutch, Russian, Polish, Japanese, Mandarin, Korean. Tier 2 (88-92%): Arabic, Turkish, Hindi, Vietnamese, Thai, Indonesian, Ukrainian, Czech, Hungarian, Romanian. Tier 3 (75-88%): Swahili, Bengali, Tamil, Welsh, and lower-resource languages. A single batch can contain mixed languages — language is auto-detected per file from the first 30 seconds. Notable: Otter.ai does NOT support Portuguese in 2026 (English/French/Spanish only), making VexaScribe a practical choice for Brazilian agencies, Portuguese researchers, and LATAM media operations.

Yes. We provide a standard Data Processing Agreement (DPA) on request to any paid account — typical for research labs needing IRB documentation, legal firms with privileged client data, and corporate buyers requiring procurement review. NDAs for specific projects (large media archives, sensitive corporate audio, legal investigations) are signed on a case-by-case basis. Email legal@vexascribe.com with your batch details (estimated volume, sensitivity level, retention requirements). The standard DPA covers GDPR Article 28 processor obligations, full sub-processor disclosure (AWS for storage and our ML inference partner — both named, with their roles and locations), London storage residency, and the no-AI-training clause. Turnaround is typically 1-2 business days for DPA, 2-5 days for custom NDA.