Self-hosted Otter alternative: a transcription web app for your team on your own GPU

A self-hosted alternative to Otter for a team

A team that uploads recordings to Otter, Happy Scribe or a similar cloud service can have the same routine on its own hardware: a web app on a workstation in the office takes the audio file, writes the transcript with speaker labels, drafts a summary with a local language model and exports the text. Scriberr and Whishper are two open web apps for this as of October 2026, and one NVIDIA RTX PRO 4500 Blackwell with 32 GB of memory runs either of them for a team. What you give up is mainly the meeting bot that joins calls by itself, and the mobile apps.

This comparison is for office managers, team leads and IT administrators who want to keep interviews, customer calls and internal meetings away from an outside service. A few terms first. A transcript is the written text of a recording. Speaker labels, which engineers call diarisation, mark each passage as speaker 1, speaker 2 and so on. A speech recognition model is the AI that turns audio into text; Whisper, Parakeet and Canary are the common open ones. The GPU, or graphics card, is the part of the computer that runs these models quickly. The sizing arithmetic for larger servers is in our guide to Whisper GPU requirements; here the subject is the app your team works in every day.

What the cloud service does and what a local app replaces

Otter describes its product on its home page, read on 30 October 2026, as an "Otter AI Meeting Agent" that "supports real-time transcription, live chat, automated summaries, insights, and action items". Through Google Calendar the agent can be scheduled to join Zoom, Teams and Google Meet meetings automatically, and the desktop app records computer audio "without bots joining the call". The page links iOS and Android apps and lists integrations with Slack, Salesforce, HubSpot, Notion and others. It states that Otter "is built with the security, governance, and compliance controls enterprises expect".

Happy Scribe, the other service teams often compare, describes on its home page, read the same day, AI transcription of audio and video "in 150+ languages", human transcription, subtitles, a calendar notetaker or recording "with no bot joining the meeting", and export to DOCX, PDF, SRT and VTT. It states that recordings are "stored in the EU". A local app goes one step further for a team that wants no outside processor at all: the recordings stay on the company's own machine.

Otter home page; Scriberr and Whishper READMEs on GitHub; all read on 30 October 2026.

With a local app, the audio file, the transcript, the speaker labels and the summary stay on the workstation, provided the summary model runs there too. Scriberr talks to "Ollama or OpenAI API compatible providers", and its release v1.2.0 added OpenAI transcription support, so the person who configures it decides whether audio and transcripts stay on the machine or go to a hosted service. For a local set-up, transcription runs on the workstation's own models and summaries go to Ollama or vLLM on the same machine. A meeting held on Zoom or Teams still passes through that service, and the local app keeps only the transcription and the summary in-house. The usual way into the app is a recording made on a participant's computer or a dictaphone file, uploaded after the meeting.

Scriberr, Whishper, WhisperX and Speaches compared

Four open projects come up when teams search for a self-hosted transcription tool. Two are finished web apps for end users; two are building blocks for an IT team that wants its own pipeline.

GitHub pages of Scriberr, Whishper and Speaches, PyPI page of WhisperX; read on 30 October 2026. Stars on GitHub: Scriberr 3.1k, Whishper 3.1k, Speaches 3.7k.

Scriberr calls itself "an open-source, and completely offline audio transcription application". It runs NVIDIA's Parakeet and Canary models and the Whisper models with word-level timing, detects speakers automatically, and lets users ask questions about a transcript or have it summarised through a connected language model. The transcript viewer follows the audio during playback and jumps to the place a user clicks in the text; users can mark passages and add notes. A folder watcher picks up new files, and its APIs connect other systems. The latest release on GitHub, v1.2.0, adds webhook notifications and a command-line tool. The README does not list export formats or user roles; its configuration includes a secret for login tokens, so plan one shared workspace per team behind your company network.

Whishper describes itself as a way to "transcribe any audio to text, translate and edit subtitles 100% locally with a web UI". It uses faster-whisper, takes uploaded audio and video files, translates through a bundled LibreTranslate service, and has a subtitle editor with segment splitting, warnings for too many characters per second and highlighting synchronised with playback. It exports TXT, JSON, VTT and SRT. User authentication is still an open item on its roadmap, so the app belongs on a trusted internal network or behind a reverse proxy that handles the login.

WhisperX is a Python tool for IT teams. Its PyPI page claims "batched inference for 70x realtime transcription using whisper large-v2", states that its faster-whisper backend "requires <8GB gpu memory for large-v2 with beam_size=5", and adds word-level timestamps through wav2vec2 alignment, with default alignment models for English, French, German, Spanish and Italian; the French, German, Spanish and Italian defaults are torchaudio's VoxPopuli models under CC BY-NC 4.0, a non-commercial licence, so a company set-up passes an alignment model whose licence allows commercial use. Speaker labels come from pyannote and need a Hugging Face access token plus acceptance of the conditions of the community-1 model.

Speaches is an "OpenAI API-compatible server supporting streaming transcription, translation, and speech generation", built on faster-whisper. It suits a company that wants one central speech service which other tools call, such as the dictation apps in our guide to local AI dictation as a Dragon alternative. Its latest release on GitHub, v0.9.0-rc.3, is a release candidate, so pin the version you deploy. A researcher who transcribes interviews alone on one computer is better served by the desktop app in our guide to interview transcription with noScribe.

Maintenance status and licences for company use

Both web apps come from small open-source projects, and both say so. Scriberr's README has an undated status section: it says development "is paused for now, but definitely not abandoned", and a summary line at its top announces "I'll be starting development again this week". The latest release on GitHub is still v1.2.0, so check the commit history before you install. Whishper's README says the project is being rewritten on a version 4 branch and that the current main branch "will not receive any new releases or updates". Neither means the software stops working. It does mean that a company should pin the version it installs, keep the container image and the model files on its own storage, name the person who reads release notes and applies security fixes, and treat the app as replaceable. Transcripts exported in open formats outlive any change of app.

The licences allow company use, with details to note. Scriberr is under MIT. Whishper is under the GNU Affero General Public License version 3; section 13 of its LICENSE file requires that a modified version offers its source code to the users who work with it over a network. Running it unchanged inside the company is the common case; if you change the code, your legal counsel reads that section first. OpenAI releases Whisper's code and weights under the MIT licence. Scriberr's README names a managed WhisperX environment for its models, so the alignment licence noted under WhisperX above applies here too: in a company set-up we check which alignment model the app loads for German and use one whose licence allows commercial use. Parakeet TDT 0.6B v3 and Canary-1B v2 are under CC BY 4.0 and their cards state they are ready for commercial and non-commercial use. The pyannote speaker-diarization-community-1 pipeline is under CC BY 4.0 and gated: you accept its conditions on Hugging Face and use an access token, after which a local copy "works without internet connection". The same card offers a precision-2 model that runs on pyannoteAI's servers; a local set-up leaves it out.

RTX PRO Blackwell cards and the right container image

Scriberr ships its Docker images in three variants: one for CPUs, one with CUDA for Pascal to Ada GPUs, and scriberr-cuda-blackwell for Blackwell. The README is explicit for RTX 50-series cards: "You must use the Blackwell-specific image", because the standard CUDA image does not work with them for reasons of PyTorch CUDA compatibility. It does not name the RTX PRO cards, but its table assigns that image to sm_120, the compute capability of the RTX PRO Blackwell cards, so the same image applies; the first test after installation confirms it.

The other steps are similar for every app. The NVIDIA Container Toolkit has to be installed for GPU access, the first start downloads the PyAnnote models, which can take several minutes, the gated pyannote model needs a Hugging Face token, and a machine that should later run without internet access needs all models stored locally. None of this is difficult for an administrator, but it is where a first installation usually stalls. You can order the workstation alone, the setup of Scriberr or Whishper with the speech models and a local summary model 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.

Which workstation for which team

Transcription itself is a light load for a GPU. On one A100 in the Open ASR Leaderboard paper (arXiv 2510.06961v3, December 2025), at batch size 64 where memory allowed, Whisper large-v3 transcribed 145.5 seconds of audio per second of computing and Parakeet TDT 0.6B v3 about 3,333. By our arithmetic on that card, a one-hour recording takes about 25 seconds with Whisper and about one second with Parakeet. We found no published figure of this kind for the RTX PRO cards, and the pyannote card gives no speed for speaker labels, which add time to every job. The card choice therefore depends on two other questions: whether the summary model runs next to the speech model, and how many uploads are processed at the same time.

Usable memory by our users-per-card rule (90 per cent of the memory the driver reports, less 3 GiB): 18.5, 25.7, 40.0 or 61.5 GiB on the workstation cards; Whisper large about 10 GB (OpenAI's Whisper repository), gpt-oss-20b 13.76 GB of weights (Hugging Face files page); our estimate.

The arithmetic behind the second row: Whisper large at about 10 GB is 9.3 GiB and gpt-oss-20b, an open model under Apache 2.0, has 13.76 GB of weights, 12.8 GiB. Together that is 22.1 GiB, more than the 18.5 GiB our rule leaves on a 24 GB card and within the 25.7 GiB on the 32 GB RTX PRO 4500, with about 3.6 GiB left for the summary model's working memory, by our estimate. The speaker-label model comes on top; its card states no memory figure, and our guide to speaker diarization on-premise gives NVIDIA's GPU-time figures for the Nemotron diariser, while pyannote publishes none. On the RTX PRO 4000 the app can still transcribe and label speakers, and summaries run when the speech model is unloaded or on a separate machine. The 48 GB RTX PRO 5000 leaves about 18 GiB free after both models, enough for parallel jobs from several departments. The RTX PRO 4500 is rated at 200 W, so a one-card workstation stays far below the 2.94 kW that a 230 V 16 A circuit carries at 80 per cent load.

When transcription becomes one service among several on the company AI server, the entry server on our AI servers page fits: two NVIDIA RTX PRO 6000 Blackwell Server Edition cards with 96 GB each, about 1.8 kW at the wall by our estimate. Our guide to meeting minutes with a local AI shows that layout, with speech models in MIG instances of one card and gpt-oss-120b on the other. For a single workstation shared by many staff, our comparison of a GPU server or workstations for a team helps with the choice.

What the team should check before switching

A short trial with your own recordings answers most questions before anything is bought. Take 20 to 30 minutes of typical audio, such as a customer call, an internal meeting and a voice memo, and run it through both apps.

• Languages. Check the accuracy in the languages you work in; our comparison of German speech recognition models and dialects gives the published figures for German.
• How audio arrives. Neither app joins video calls by itself. Decide who records and where files are uploaded, or use Scriberr's folder watcher on a shared folder; our guide to an AI voice recorder without the cloud describes that route for dictaphone and phone recordings.
• Exports. Whishper exports TXT, JSON, VTT and SRT; Scriberr's README lists no formats. If transcripts must end up as Word documents, plan a conversion step.
• Access. Put the app behind your company login or VPN. Whishper has no user authentication yet.
• Informing and retention. Tell participants before a meeting or call is recorded and record only when nobody objects, and decide when audio and transcripts are deleted; our meeting-minutes guide covers both under the GDPR.
• No outbound calls. Check that summaries go to the local model and that no component sends audio to a hosted service, as our article on telemetry in self-hosted AI tools describes. Our article on local AI and the GDPR explains which duties stay with you.

Send us your languages, your hours of audio a week and the number of users, and we suggest the best solution for your tasks and budget.

FAQ