AI server for media: transcripts, subtitles, translation and archive search

What an AI server does for broadcasters and publishers

An AI server for media turns programmes, interviews, podcasts and archive recordings into transcripts, subtitle files, translations and a searchable archive, on hardware the broadcaster or publisher controls. Speech recognition needs little GPU memory, so the server is sized by hours of audio and live channels; one with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards runs recognition, speaker labels, translation, a summary model and the search index side by side. Four or eight cards take a large archive project next to daily production.

The figures in this guide are published measurements, with their source and conditions, or our estimates, labelled as such. Our detailed guides on Whisper GPU requirements and on-premise machine translation carry the full sizing of the two largest parts.

The task in plain words

For a director of a broadcaster, a publisher or a production company, three numbers describe the job. The first is the hours of new material per day: news, shows, interviews, podcasts and video for the website. The second is the archive, the hours already digitised and waiting to be transcribed once. The third is the number of live channels that need subtitles at the same moment. The languages come on top, because they decide the models.

Daily production and live subtitles are a steady, modest load. The archive is a one-off peak: tens of thousands of hours that the server works through for days or weeks, in the background, while daily work goes on. Once transcribed, the archive becomes searchable in plain language, so an editor who asks for every interview in which a minister spoke about rail freight gets the programmes and the minute in each.

Transcripts: models, languages and speed

OpenAI's Whisper repository (read 24 October 2026) and Hugging Face model cards; NIM profile from NVIDIA's ASR NIM support matrix (7 October 2026); RTFx from the Open ASR Leaderboard paper (arXiv 2510.06961v3, 10 December 2025), one A100-SXM4-80GB, batch 64 where memory allowed; as quoted in our Whisper GPU guide.

RTFx is the inverse of the real-time factor: seconds of audio transcribed per second of computing. By our arithmetic, a 10,000-hour archive takes about 69 hours of computing with Whisper large-v3 on the paper's A100 and about 3 hours with Parakeet TDT 0.6B v3. OpenAI's repository says the turbo model runs about eight times as fast as large. On an L4, by the estimates in our Whisper guide from E2E Networks' test, the same archive takes about 42 hours with Parakeet and about 61 hours on four cards with Whisper turbo. NVIDIA's model card lists the L4 as test hardware for Parakeet TDT 0.6B v3 but does not name its Ada Lovelace architecture as supported.

Accuracy decides between them as much as speed. For German, the paper's multilingual table gives a word error rate of 4.97 per cent for Whisper large-v3, 4.90 for Parakeet TDT 0.6B v3 and 4.96 for Canary-1B v2. That table covers German, French, Italian, Spanish and Portuguese only, so for Polish, Czech or Romanian test the models on a few hours of your own programmes first. Broadcast audio with music beds, studio guests talking over each other and phone interviews is harder than the read speech of the benchmarks.

The published figure for the card we build on comes from MLPerf. In MLPerf Inference v6.1, entry 6.1-0050, an HPE ProLiant Compute DL380a Gen12 with eight RTX PRO 6000 Server Edition reached 19,278.3 tokens per second on the Whisper benchmark in the offline scenario, about 2,410 per card by our division. MLCommons' description says the benchmark runs Whisper-Large-V3 on LibriSpeech clips of up to 30 seconds and counts tokens of transcript per second, so the figure ranks systems but does not convert directly into hours of your audio.

MLPerf® Inference: Datacenter v6.1, closed division, available category, Whisper, offline scenario, entry 6.1-0050, retrieved from mlcommons.org on 24 October 2026. Result verified by MLCommons Association. Per card is our division of the system result, not an MLPerf metric. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

Subtitles from the transcript

Whisper's command-line tool writes subtitle files itself: --output_format accepts txt, vtt, srt, tsv and json, among others. Its --max_line_width and --max_line_count options shape the lines to a house style, but both require --word_timestamps True, which the tool marks as experimental. Parakeet TDT 0.6B v3 returns punctuation and word timestamps, from which the pipeline builds subtitle lines. In either case a subtitle editor reviews timing and line breaks before broadcast.

Speaker changes need diarisation, and it can cost more than the transcription. In NVIDIA's L40S figures, Parakeet 0.6B CTC offline reached an RTFx of 3,638 without diarisation and 101.5 with it. NVIDIA's Nemotron 3 Diarization, published under OpenMDW 1.1 with 100M parameters and up to eight speakers, reports an RTFx of about 15,113 at batch 32 on an RTX PRO 5000 with a compiled model. For a talk show with several guests, size from the diarised rate.

Live subtitles need a streaming model, because NVIDIA's ASR NIM runs Whisper offline only. NVIDIA's figures of 7 October 2026 give one L40S a maximum of 190 live English streams with Parakeet 0.6B CTC at 160 ms chunks. Nemotron 3.5 ASR Streaming 0.6B lists German among its 19 locales for "highest-accuracy ASR" with chunks of 80 to 1,120 ms, but we found no stream figures for it on the cards we supply. A broadcaster with a handful of live channels is a small load next to those English figures; test the model on your own live audio for latency before you rely on it.

We build speech servers for media companies; send us your hours per day, archive size and live channels, and our team replies with a configuration.

Translation of subtitles and articles

Whisper can translate speech, but its README says the translate task renders non-English speech into English, and that the turbo model "will return the original language even if --task translate is specified". Canary-1B v2 translates speech between English and the other 24 of its languages. For translation between two non-English languages, and for articles and subtitle files, a text translation model works from the transcript.

Our translation guide plans with 2 output tokens per word: one million words a day average 69 tokens per second over eight hours, and by its estimate one 48 GB card with a 9B to 12B translation LLM in BF16 carries that. EuroLLM covers all 24 official EU languages under Apache 2.0, and the TranslateGemma cards name 55 languages under the Gemma Terms of Use; the NLLB-200 cards call it a research model, not released for production deployment. Subtitle lines are short, so send a whole scene per request: the model keeps names and terms consistent with more context. Argos Translate's file library lists .srt among its supported formats.

For newsrooms one rule matters in particular. Since 2 August 2026, Article 50 of the AI Act has required deployers to disclose AI-generated text published to inform the public on matters of public interest, unless the text has undergone human review or editorial control and someone holds editorial responsibility for its publication. Deep fakes must be disclosed as well. Our article on the EU AI Act gives general information; national practice may differ and your legal department decides.

Archive search on transcripts

Archive search turns each transcript into chunks with their timestamps, embeds the chunks as vectors and finds the nearest ones to a question, so a result opens the programme at the right minute. Metadata such as date, channel, rights and speakers filters the results.

NVIDIA NeMo Retriever embedding NIM (version 1.14.0, 30 September 2026) and reranking NIM (17 August 2026) performance pages, FP8, as quoted in our embedding and reranker guide; vector size from our vector database comparison. Archive example: 10,000 hours in one-minute chunks.

The example assumes one chunk per minute of speech that stays under 512 tokens, an assumption to check with your own transcripts. Embedding takes little time compared with transcription, and only new material is embedded after the first run. Qwen3-Embedding and bge-m3 state support for more than 100 languages, which suits archives in several languages, and Qwen3-VL-Embedding, published in January 2026, also embeds images, so keyframes of a programme can join the index; NVIDIA publishes no speed for it. Our guides to embedding and reranker servers for RAG and to RAG on company data cover models, chunking and access rights. A language model on the same server writes summaries and keywords for every programme; gpt-oss-120b fits one 96 GB card.

Configuration options by card count

The options follow the reference configurations on our AI servers page and NVIDIA's guide for certified systems: two processors, at least six cores per GPU and system memory of twice the GPU memory.

Reference configurations on our AI servers page; MIG layout as in our MIG profiles guide; power by our estimate (cards at 600 W, two AMD EPYC 9135, 9355 or 9555 for two, four or eight cards, plus 15 per cent for the rest of the server).

Two cards. The entry server, two RTX PRO 6000 Server Edition with two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives, splits the first card into four isolated 24 GB MIG instances: Whisper large-v3 at about 10 GB or the Parakeet NIM profile at 14.02 GB, the diariser, a streaming model for live subtitles and the embedding model with its reranker. The second card runs gpt-oss-120b for summaries, keywords and translation; by our users-per-card rule it holds 19 requests of 32K tokens at once.

Four cards. The inference node with 768 GB and four NVMe drives adds room for an archive project: the speech card works through the archive at night and on weekends, a second card carries search and a translation model, and two cards run gpt-oss-120b as a pair for a newsroom assistant, 187 conversations of 16K at once by our rule. During a large digitisation project, more cards can run recognition and return to their usual roles afterwards.

Eight cards or two servers. A media company with several brands and a large archive transcribes on several cards while production runs on the rest; the MLPerf entry above is such an eight-card system. Two four-card servers instead of one eight-card server keep live subtitles running while one of them is updated, as our guide to one 8-GPU server or two 4-GPU servers discusses. For transcription alone, our Whisper guide shows that one L4 or L40S in a server you already run is enough for many workloads.

The RTX PRO 6000 Max-Q is the workstation card, a 300 W blower card; in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition. We build each of these servers to order; describe your languages and the share of live and archive work, and we size the cards for it.

Notes for the IT team

1. Engines. faster-whisper is under MIT and needs no NVIDIA AI Enterprise licence. NVIDIA's ASR NIM needs compute capability 8.0, at least 16 GB of GPU memory and NVIDIA AI Enterprise in production; its support matrix lists "Blackwell RTX 60xx", so check the matrix of the release you deploy against the card.
2. MIG. Up to four 24 GB instances per RTX PRO 6000. NVIDIA's prerequisites are Linux, driver 575.51.03 or later, a current vBIOS and the DisplayModeSelector set to compute mode.
3. Video files. NVIDIA's page lists four video encode and four video decode engines on the Server Edition, which a pipeline can use to decode programme files, extract keyframes or make proxies on the same card.
4. Data protection. Recordings in which people can be identified are personal data under the GDPR. Check that no component sends audio to a hosted service, and keep recordings and transcripts on the server under your control, as our Whisper guide explains.
5. Network and updates. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere, while routes such as /invocations and /metrics answer without the key (vLLM 0.32.0 security guide), so keep the engines behind an internal gateway. Keep the NVIDIA Container Toolkit at 1.17.8 or later and the driver at a release NVIDIA lists as fixed in its bulletin of 30 September 2026. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime.

Power and room

By our estimate, the two-card server draws about 1.8 kW at the wall with two EPYC 9135 processors and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. The four-card server with two EPYC 9355 draws about 3.4 kW with the cards at 600 W, or about 2.7 kW at the 450 W cap, and needs 32 A or three-phase. Eight cards with two EPYC 9555 draw about 6.3 kW by our estimate; Dell's datasheet gives about 8 kW per eight-card node as an indicative figure, and feeds should be planned on the maker's figure. Our guide to GPU server power on 16 A, 32 A and three-phase feeds works through each case.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build the servers to order with two, four or eight RTX PRO 6000 Server Edition cards, or with L4 and L40S cards for transcription alone, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. NVIDIA AI Enterprise licences for NIM come on the same invoice. Our Private AI/ML service, delivered with our engineering partner Vixen.UNO, deploys the speech, translation and language models and the archive search on top. Our guides for e-commerce and creative agencies cover neighbouring workloads.

FAQ