Customer support · Caption Desk

Help when
you need it.

How it works

Every recording travels through the same private pipeline on the service's own servers. Drag the diagram to rotate it and click a step for details.

  1. Upload — validated before bytes are sent.
  2. Queue — in order, with operator priority.
  3. Device — NVIDIA CUDA, AMD Radeon (Vulkan) or CPU.
  4. Transcribe — timed captions with whisper.cpp.
  5. Translate — local translation models.
  6. Voices — natural neural voice-over.
  7. Video — multi-track MP4 and burned captions.
  8. Download — until the retention deadline.

First job walkthrough

Services, uploads, live progress and deliverables in seven steps.

Open the tutorial ↗

Common questions

Limits, privacy, voices, caption types, languages and expiry.

Read the FAQ ↗

Running the service?

GPUs, monitoring, the advisor, users, storage and voices.

Operator guide ↓

Services & files

ServiceYou receiveBest for
TranscriptTXTNotes, articles, search
Subtitle packageTXT, SRT, VTT, ZIPEditing in your video tool
Reviewed captionsOperator-checked TXT, SRT, VTTPublishing English captions
Multilingual videoMP4 with caption + audio tracks, per-language SRT/VTT, voice previews, optional burned MP4Reaching viewers in other languages

Accepted inputs: WAV, MP3, FLAC, OGG/Opus, M4A, MP4, MOV, MKV, WebM. Current limits: 512 MiB and 60 minutes per file, 10 free jobs per day.

Live progress & statuses

After the upload finishes, the job card shows what is happening right now and how far along it is. The page refreshes itself every few seconds.

1. Preparing audioConverted to 16 kHz speech audio.
2. Transcribing speechFollows the engine percent-by-percent.
3. TranslatingPer target language.
4. Generating voicesNatural voice per language.
5. Building MP4Tracks are muxed.
6. Burning captionsOnly if requested.
7. PackagingZIP and checks.
  • Uploading → Awaiting payment (paid modes) → Queued → Processing → Awaiting review (reviewed service) → Ready.
  • Side exits: Failed (operator can retry), Cancelled, Expired / Deleted.

Languages & natural voices

Video jobs can add translated captions and a voice-over for up to 10 target languages. Enabled languages: en,es,fr,de,pt,it,ru,ja,zh,ar,hi,nl,pl,tr,id,vi,ko,sv,cs,el,uk.

How the voice-over is timed

  1. Each caption gets a window that lasts until the next caption starts.
  2. The voice reads the translated caption at a natural pace.
  3. If a sentence is too long, it is re-spoken slightly faster by the voice itself — never stretched into a "chipmunk" sound.
  4. Lines never overlap; anything still too long fades out gracefully.
  5. The original soundtrack can remain quietly underneath and dips automatically while the voice speaks.

Good to know

  • Voices are general-purpose and never clone the speaker.
  • Translation and voice quality vary by language — always review.
  • Burned captions are permanent; selectable tracks can be turned off.
  • Use VLC for multi-audio MP4 files.

Workspaces & teams

Every account has a private workspace. Create more workspaces for teams; members share that workspace's jobs. Owners and managers add registered users and remove members. Switch workspace from the selector in My workspace.

  • Each workspace sees only its own jobs.
  • Owners and managers manage members.
  • Service operators can see all jobs to run the service.

Downloads & players

  • ZIP — everything for the job.
  • SRT/VTT — import into Premiere, Resolve, Final Cut, YouTube or a web player.
  • MP4 — original video with selectable caption and audio tracks.
  • PREVIEW-xx — a browser-friendly video per voice language.
  • BURN-xx — captions drawn permanently into the picture.

Troubleshooting

The upload is rejected immediately

The preflight check explains why: unsupported file type, file too large, language not enabled, daily or storage quota reached, or the queue is full. Fix the option and try again — nothing was uploaded.

The upload stops moving

If no bytes move for two minutes the transfer is cancelled with a network message. Retry on a stable connection or split very long recordings.

The job stays "Queued"

Other jobs are ahead of yours, or processing is paused. Progress starts as soon as a worker is free.

A language I need is missing

Ask the operator to enable it; translation and voice models are installed per language.

I can't sign in

Use password reset if email is configured. A suspended or blocked account shows the reason at sign-in.

For service operators

Operator guide

The Admin dashboard has seven tabs: Job queue, System & GPUs, Advisor, Users & storage, Voices, Service settings and Organizations.

GPUs, devices and automatic fallback

Caption Desk runs on NVIDIA GPUs (RTX 4090, RTX 4000 / RTX 4000 Ada, Tesla P4 and other CUDA cards), AMD Radeon GPUs through Vulkan, or the CPU. The worker tests every device with a real transcription before using it. Each job goes to the fastest verified device; if that GPU fails, the same job is retried on the next device and the attempt is recorded on the job card.

  • NVIDIA CUDA (preferred) → AMD Radeon Vulkan → CPU.
  • A failure on a GPU retries the job on the CPU automatically.
SettingWhat it does
Processing deviceAutomatic (recommended), GPU only, CPU only, or one specific detected device.
Parallel jobsWorker slots. Set it to the number of GPUs: each slot prefers its own GPU, so an RTX 4090 and a Tesla P4 work on two jobs at once.
Disabled devicesExclude a device (for example a GPU reserved for other work) without uninstalling anything.
CPU fallbackRetry on the CPU when a GPU fails or none is available.
Video encoderNVENC (NVIDIA) or VA-API (AMD) for burned captions; libx264 is always the final fallback.

The Job queue shows the device a running job is using; finished jobs show Converted on … with “GPU confirmed by engine log”, any fallback, encoders and voices used. Use Stop to end a running job, Prioritize to move a queued job up, and Delete files to free space.

Hardware, capacity and launch guide

Administrators find a full guide at Admin → Hardware & launch guide: what works on NVIDIA, AMD Radeon, Intel and CPU-only servers; how much video memory each Whisper model needs and how many jobs fit on a 4, 8, 16 or 24 GiB card; why translation, voices and software video encoding depend on CPU cores; how much RAM and disk each conversion slot uses; a live summary of this server with a suggested number of slots; how many users the site serves and how the fair queue handles many uploads at once; and a launch checklist to finish before posting the link publicly.

System monitor and advisor

System & GPUs shows live CPU, memory, swap, disk space, disk and network throughput, and for every GPU its utilization, VRAM, temperature, power and fan, with 30-minute and 24-hour charts. It also lists the verified processing devices, what each worker slot is doing, and which video encoders passed their test. Press Re-detect devices after driver or hardware changes.

The Advisor re-checks the server every few seconds: OS support, pending reboots and updates, time sync, CPU governor and threads, memory and swap, disk and backups, GPU drivers and engines (including Tesla P4 driver limits and passive cooling), model size versus VRAM, hardware encoders, voices, translation models, worker health, failures and security. Each finding has a recommendation and, where useful, a command you can copy.

Users and storage

  • Add user with a password or a generated temporary password (shown once).
  • Suspend for a number of days or indefinitely, with a reason shown at sign-in; it lifts automatically on the end date.
  • Block permanently; the email cannot be re-registered.
  • Reactivate, Make / remove administrator (the last administrator is protected), Reset password, Sign out everywhere.
  • Storage quota per account — uploads beyond it are refused.
  • Delete account removes its jobs and files; shared workspaces are handed to another member.

The disk panel measures real usage per account and per category (customer jobs, translation models, voices, backups, database), shows the whole disk and lists orphaned folders.

Voices

Install the recommended natural voice for each enabled language with one click, pick alternative voices from the catalog, and press Play to hear a sample with your own text. Downloads are verified by size and checksum. Adjust speaking rate, soundtrack level and maximum speed-up under Service settings → Voices & dubbing. Voice licences differ per model; review them before commercial use.

Acceptance test

After installing, and after any change to GPUs, drivers, engines or voices, run the end-to-end self-test. It transcribes a sample on every GPU and the CPU (a GPU that silently falls back to the CPU fails), encodes video with every encoder, burns in subtitles, speaks with every installed voice, translates with every installed language pair and checks GPU monitoring and the database. It changes nothing except storing its result, which the Advisor tab shows.

Server commands

sudo python3 install_captiondesk_unified_v26.py --acceptance-test
systemctl status captiondesk-web captiondesk-worker caddy --no-pager
journalctl -u captiondesk-worker -n 100 --no-pager
sudo runuser -u captiondesk -- env CAPTIONDESK_CONFIG=/etc/captiondesk/config.json HOME=/var/lib/captiondesk /opt/captiondesk/current/.venv/bin/python -m captiondesk.worker --detect
sudo python3 install_captiondesk_unified_v26.py --upgrade-only --install-voices
sudo python3 install_captiondesk_unified_v26.py --backend auto --model large-v3-turbo
sudo python3 install_captiondesk_unified_v26.py --reset-admin-only --admin-email you@example.com

Forgotten administrator password, backups and restore

On the server, sudo captiondesk-admin reset-password finds the administrator account, asks for a new password twice, makes a backup first, clears login lockouts and any block, signs out old sessions and prints the address and email to sign in with. unlock clears lockouts without changing the password. Backups run daily; set Service settings → Off-server backups to copy them to a NAS or another server (run backup-key first and add the printed key there). restore checks a backup, keeps the current database and puts it back automatically if the service does not start.

sudo captiondesk-admin reset-password
sudo captiondesk-admin list-admins
sudo captiondesk-admin unlock
sudo captiondesk-admin backup-now
sudo captiondesk-admin backup-key
sudo captiondesk-admin restore /var/lib/captiondesk/backups/captiondesk-YYYYMMDD-HHMMSS.zip
sudo captiondesk-admin orphans
sudo captiondesk-admin restore-media

Site analytics

Admin → Analytics shows visitors, page views, where visitors come from (search engines, AI assistants such as ChatGPT and Perplexity, social networks, other sites, utm_source tags), landing pages, countries (behind Cloudflare), devices, browsers, pages not found, crawlers by name, and registrations, sign-ins, failed sign-ins and uploads, for 7 to 365 days with a CSV download. No cookies are used and visitor addresses are never stored; your own administrator visits are not counted. Switch it off, choose the time zone for its days, or change how long daily counts are kept in Service settings → Site analytics.

Search engines, social networks and AI assistants

Public pages carry titles, descriptions, link previews and structured data automatically, with a site map at /sitemap.xml and a summary for AI assistants at /llms.txt. In Service settings → Search, social and AI add your Google and Bing verification codes (then submit the site map in both consoles), your share image and social profiles, and choose whether AI crawlers may read the site. Bing, Yandex, Seznam and Naver are notified of page changes automatically through IndexNow (or press Notify search engines now); Google reads the site map. The site is only indexed when it is served over https.

Contact the operator

Email info@tdemo.com for job errors, account access, billing concerns or deletion requests. Include the job ID; never send your password.