Opus Clip alternative: shorts from webinars and podcasts with AI on your own workstation

Shorts from long videos without uploading them

A local alternative to Opus Clip turns a webinar, a podcast or a conference talk into vertical shorts on your own workstation: WhisperX transcribes the recording with a timing for every word, a language model on the same machine reads the transcript and proposes the best passages, face tracking turns the 16:9 picture into a 9:16 frame that follows the speaker, and the captions are rendered into the clip. Open projects such as OpenShorts and ClipsAI cover some or all of these steps, and the chain can also be built from individual open tools, which is the setup we offer. By our users-per-card rule, one NVIDIA RTX PRO 5000 Blackwell with 48 GB holds the speech and language models together.

The job is routine in marketing. A company records a monthly webinar, a podcast with a customer or a talk at a trade fair, and the social media team wants eight to ten clips of 30 to 60 seconds from each. Cutting them by hand means watching the hour, picking the moments, cropping, adding captions and exporting each clip. Cloud tools automate these steps and need the recording on their servers. This comparison is for marketing teams, agencies and in-house video editors who want the same workflow on their own machine, and for the IT colleague who sets it up.

Reframing means cutting a vertical frame out of a horizontal picture and moving it with the person who is speaking. Burned-in captions are subtitles drawn into the picture itself, so they show on every platform without a separate subtitle file.

Opus Clip and the open options side by side

opus.pro feature pages; GitHub READMEs of OpenShorts, ClipsAI, AI-Youtube-Shorts-Generator and FunClip, and the OpenShorts licence file; ClipsAI documentation and PyPI release history; all read on 30 October 2026.

OpusClip. Its pages, read on 30 October 2026, describe ClipAnything, which "turns any genre" of video "into viral clips in 1 click", ReframeAnything, which "resizes any video for any platform" with AI object tracking, captions added "with over 99% accuracy" by the provider's statement, AI B-roll, brand templates with font, colour, logo and intro, a scheduler for social platforms, team accounts and an API. On its clip-maker page, "AI determines the virality potential of each clip". Input is a file upload or a link; for its Pro plan the page lists YouTube, Zoom, Vimeo, Google Drive and other sources, and the provider names English, German, Spanish, French, Portuguese "and other 20 more languages". It covers the route from upload to scheduled post, and the recording is processed by the provider's service.

OpenShorts. An "Open source AI clip generator" with 7.1k stars on GitHub on 30 October 2026. Its licence file is MIT, and its README places only the cloud/ directory, with billing and the hosted service, under a separate commercial licence; self-hosting the core app "never requires this directory". Transcription uses faster-whisper with word timings, ffmpeg cuts, its tracking mode follows faces with MediaPipe and YOLOv8, and the subtitles are burned in locally. The step that picks the moments sends the transcript, never the video, to Gemini by default; the README describes pointing it at "Ollama, LM Studio, vLLM or any OpenAI-compatible server" instead, after which "the whole pipeline stays on your box". Features that look at the frames, such as its automatic choice of layout, need a Gemini key and otherwise fall back to the plain face-tracking crop; dubbing and publishing use outside services. For YOLOv8 its tech stack names the ultralytics package, which is under the AGPL-3.0 and whose maker directs commercial users to an enterprise licence. For that reason OpenShorts is not part of the setup we offer; have your legal counsel check the licences before you use it.

ClipsAI. "Clips AI is an open-source Python library that automatically converts long videos into clips", written for "audio-centric, narrative videos" such as podcasts, interviews and speeches, under the MIT licence. It finds clips by analysing the WhisperX transcript and reframes to 9:16 "to focus on the current speaker". Its documentation names the parts: pyannote's speaker-diarization-3.0 to tell who speaks when, PySceneDetect for scene changes, and MTCNN and MediaPipe for faces. The resizer needs a Hugging Face access token for the pyannote model. The latest release on PyPI is 0.2.1 of 15 January 2024.

AI-Youtube-Shorts-Generator. An MIT project with 5.6k stars on 30 October 2026, last updated on 22 October 2026. Its local mode transcribes with faster-whisper on CPU or CUDA and crops with ffmpeg and OpenCV face tracking, while the LLM step needs an OpenAI or Gemini key; its API mode sends download, transcription, ranking and cropping to a hosted clipping service. In the documented setup the transcript therefore leaves the company.

FunClip. A "FunASR-powered video transcription, subtitle generation, and LLM-assisted clipping tool with a local Gradio UI", MIT for the code, version 2.2.1 of 1 September 2026. Chinese is its primary language, its LLM options are hosted gateways with API keys, and it does not describe vertical reframing; its README names no German speech model, so for German webinars it fits least of the projects here.

What stays in the company with each option

With OpusClip, the video file or the link goes to the provider and the clips come back from its service. With AI-Youtube-Shorts-Generator in local mode, the video stays, and the transcript goes to the language model provider. OpenShorts sends the transcript to Gemini by default and keeps everything on your machine once it points to a local model server. With ClipsAI, FunClip in its local functions and the chain described below, the recording, the transcript and the clips stay on your workstation, provided the language model also runs there. The pyannote model behind speaker detection is gated on Hugging Face: you accept its conditions and download it once with a token, after which it runs on the workstation.

That matters for internal town halls, customer webinars with questions from named participants, product launches before the release date and recordings with guests who agreed to a webinar, not to a public clip. Before you publish a clip of a guest or a customer, check that their consent covers short clips on social platforms; your data-protection officer or legal counsel confirms what is needed. Clip only recordings your company owns or may edit: OpenShorts can also fetch videos by URL with yt-dlp, and someone else's video needs the rights holder's permission. Our guide to local AI and the GDPR covers what own hosting solves.

The local chain, step by step

1. Transcript with word timings. WhisperX (BSD-2-Clause, version 3.8.6 of 25 May 2026) runs Whisper with "Accurate word-level timestamps using wav2vec2 alignment"; plain Whisper timestamps, in its README's words, "can be inaccurate by several seconds". Default alignment models cover English, French, German, Spanish and Italian; for the last four they are torchaudio's VoxPopuli models under CC BY-NC 4.0, a non-commercial licence, so we set a commercially licensed alignment model with --align_model. Speaker labels come from pyannote's community-1 model. The comparison of Whisper runtimes is in our guide to faster-whisper, WhisperX and whisper.cpp.
2. Highlights. A local language model reads the transcript with sentence timings and returns a list of passages: start, end, a working title and the reason it works as a clip. You can tell it what you want: product questions, a customer quote, a clear statement in under 45 seconds.
3. Cut and reframe. ffmpeg cuts each passage, and the reframing step finds the speaker's face and moves a vertical frame with it. ClipsAI's resizer does that from speaker turns and face detection, with settings such as the sampling per speaker segment.
4. Captions. The word timings drive the captions; WhisperX's --highlight_words option writes the timing of each word into the SRT file for the word-by-word style common on shorts. The captions are then rendered into the picture with your fonts and colours.
5. Review and export. An editor looks at the proposals, trims start and end, drops the weak ones and exports. The model proposes; a person chooses what is posted.

A transcript is small for a language model. By our estimate, even a fast speaker at 200 words a minute produces 12,000 words in an hour, about 24,000 tokens at our planning figure of 2 tokens per word, less than a tenth of the 262,144-token context of Qwen3.8-27B. NVIDIA's card for its NVFP4 build lists text, image and video as inputs, so key frames can be added when a slide or a demo matters more than the words; we found no published test of highlight selection with it.

You can order the workstation alone, the setup of WhisperX, a local model and ffmpeg alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Captions in German and other languages

WhisperX has a default alignment model for German, but it is one of the non-commercial VoxPopuli models, so for client work the setup uses a commercially licensed German alignment model, set with --align_model as described above. For Dutch, Polish, Portuguese, Czech, Swedish and many other European languages, WhisperX lists alignment models from Hugging Face in its code; test the one for your language on your own recordings. WhisperX also notes that words containing characters missing from the alignment model's dictionary, typically numbers with symbols, get no timing, so check figures and product codes in the captions. Product names and company terms are best fixed in the review step and kept in a word list for the next video. Our article on German speech recognition accuracy has the published error rates for German. For a translated voice see our guide to AI video dubbing and lip sync, and for a narrator on product videos our guide to AI voice-over for marketing videos.

Workstation sizes from one marketer to a media house

Four kinds of work load the card: speech recognition, the language model, face detection and video encoding. The models are modest; the point of a larger card is keeping them loaded together and rendering several clips at once. We size memory with our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card, less the weights. OpenAI's Whisper repository gives about 10 GB of GPU memory for the large model, and WhisperX states that it "requires <8GB gpu memory for large-v2 with beam_size=5".

Our estimate by the users-per-card rule; Whisper large about 10 GB (OpenAI's Whisper repository); gpt-oss-20b 12.8 GiB, Qwen3.8-27B NVFP4 20.4 GiB and FP8 28.75 GiB as in our model guides; encoder counts from NVIDIA's product pages and its video encode and decode support matrix.

On the 48 GB RTX PRO 5000, Whisper large (about 9.3 GiB) and Qwen3.8-27B in NVFP4 (20.4 GiB) take 29.7 of the 40.0 GiB the card offers by our rule, which leaves about 10 GiB for the speaker model, face detection and the language model's work, by our estimate. The 72 GB card holds the FP8 build with about 23 GiB to spare. NVIDIA lists three ninth-generation NVENC encoders, the hardware units that encode video, and three decoders for the RTX PRO 5000, two of each for the RTX PRO 4000 and four of each for every RTX PRO 6000 edition; more encoders let several clips render at once. The RTX PRO 5000 draws up to 300 W.

We found no published speed figures for the whole chain on RTX PRO cards; the figures that exist cover single steps. pyannote states a real-time factor of about 2.5 per cent for speaker-diarization-3.0 on one Tesla V100 GPU, with an Intel Cascade Lake 6248 processor for the clustering, or about 1.5 minutes per hour of conversation. StorageReview's review of 14 October 2026 measured Qwen3.8-27B in NVFP4 on cards including the RTX PRO 4500 Server Edition in a Dell PowerEdge R770 server; with one request at a time, every card landed between 17 and 22 tokens per second. By our estimate a list of ten proposals of about 100 tokens each then takes under a minute to write at that speed, with the model's thinking mode switched off, plus the time to read the transcript. Transcription and rendering time depend on the length, the model size, the batch size and the encoder settings; a test with one of your own recordings gives the figure.

The RTX PRO 6000 Workstation Edition is a 600 W card, and two of them are the realistic maximum in one workstation; three or four RTX PRO 6000 Max-Q belong only in a chassis built for them with strong directed airflow. When the studio tier also generates B-roll or upscales footage, our guides to the AI video workstation for marketing and to local AI video upscaling size those jobs. A media house with daily hours of footage is better served by a server, as our guide to an AI server for media describes.

We build these workstations to order; send us the type and length of your recordings, the clip lengths you need and the number of videos per month, and our team replies with a configuration.

Notes for the IT team

1. Pin ClipsAI's environment. Its last release is from January 2024 and it installs WhisperX from Git; build it in its own container and test after every update of WhisperX or pyannote.
2. Hugging Face tokens. pyannote models are gated; download them once and run offline afterwards. WhisperX's README states a CC-BY-4.0 licence for the community-1 diarization model.
3. Structured output. Ask the language model for the clip list as JSON with start and end times against a fixed schema, and reject times outside the recording.
4. Encoding. Check the encoder load in nvidia-smi while several clips render, and plan the card from that measurement.
5. Storage. Raw webinars are large; keep sources, proposals and exported clips apart and delete raw files on a schedule.

Our Private AI/ML service deploys open models on-premise with vLLM or Ollama, so recordings and transcripts do not leave your network.

FAQ