AI voice recorder without the cloud: a Plaud alternative on your own workstation

A Plaud alternative that keeps client conversations in the company

An AI voice recorder without the cloud does what Plaud and similar devices do, recording a conversation, transcribing it and writing a summary, with the transcription and the summary running on a workstation in your own office. The recording comes from an ordinary voice recorder or a phone, reaches the workstation through a watched folder or a web app, and an open app such as Scriberr writes the transcript with speaker labels and drafts the notes with a local language model. One workstation with an NVIDIA RTX PRO 4000 Blackwell card serves a team, and no conversation leaves the company for processing.

This guide is for lawyers, consultants, auditors, managers and field staff such as service technicians and surveyors, who record client meetings, site visits and phone notes and cannot send them to a provider's cloud. Meetings in a conference room, with participants informed and minutes approved, are covered in our guide to meeting minutes with a local AI; our comparison of a self-hosted Otter alternative looks at transcription apps for teams, and our Gong alternative at the analysis of sales calls. This article is about the recorder in the pocket.

Some terms used below. 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 turns audio into text; Whisper from OpenAI and Parakeet from NVIDIA are the common open ones. A language model writes the summary from the transcript. A watched folder is a folder on the workstation that an app checks for new files and processes by itself.

What Plaud's own pages say about processing

Plaud describes its product on its EU home page, read on 30 October 2026, as an "AI note taker for every conversation" that transcribes and summarises recordings, with devices such as Plaud Note, Plaud Note Pro and the wearable NotePin. The processing is described on three pages, none of them dated:

• Trust centre. "All data is transferred to the cloud for processing." Plaud names data centres in four regions, among them "EU AWS (Frankfurt)", and "users in Europe will have EU-hosted, enterprise-grade subprocessors". Audio and transcripts "are not stored in the cloud unless the Cloud Sync feature is enabled by the user". The page lists "GDPR, ISO 27001, ISO 27701, SOC2 Type 2, HIPAA and EN 18031" and states that "All LLMs used in Plaud have zero training and zero data retention policies, enforced contractually via DPAs."
• Help centre on AI processing. Plaud uses three named AI providers "for transcription and summarization", "routes AI processing requests to regional endpoints based on your account's location", and with zero retention "audio, transcripts, and prompts are not stored, cached, or logged by the provider after the response is returned". "The cloud is only contacted on-demand when you explicitly request AI processing", for a transcript or a summary.
• Help centre on security. Storage for "EU / EEA" is "AWS Frankfurt (Germany)"; with Cloud Sync disabled, "Recordings stay on-device; cloud used only on-demand for AI processing".

By these pages, Plaud keeps EU data in Frankfurt and binds its AI providers to zero retention, while every transcript or summary still sends the audio or the text to Plaud's cloud and one of those providers. Many companies accept that with a data processing agreement. A law firm, an auditor or a consultancy whose clients or professional rules exclude outside processors for such conversations needs the processing on its own hardware.

The local chain: record, upload, transcribe, summarise

Scriberr README and usage guide; Ollama library page; all read 30 October 2026. The chain is our description.

Scriberr calls itself "Self-hosted AI audio transcription". It runs NVIDIA's Parakeet and Canary models and "the older more popular Whisper models", detects speakers automatically, and lets users "generate summaries, ask questions, or have a full conversation with your transcripts right inside the app". Its usage guide says users "choose the model for summarization and the template prompt from available templates", and that "Either the OpenAI or Ollama credentials must be configured as environment variables for summarization and chat to work".

That last sentence decides what stays local. If only Ollama on the workstation is configured, the audio, the transcript, the speaker labels and the summary stay on that machine. In Ollama, an open model server under the MIT licence, gpt-oss-20b is a 14 GB download; tags ending in "-cloud" show no download size, and a local set-up uses only models stored on the machine. Scriberr's release v1.2.0 added "OpenAI transcription support"; for a local set-up that option stays unused, and our article on telemetry in self-hosted AI tools shows how to check that no component calls out.

Getting recordings to the workstation without manual steps

A Plaud device passes its recordings to the Plaud app on the phone. The local chain needs its own route to the workstation; there are three, and they can run side by side.

A voice recorder and a watched folder. Scriberr's README lists a "Folder Watcher that automatically processes new files in a folder". The files of a recorder that the PC sees as a USB drive are copied to a shared folder on the company network, and the workstation picks up each file. Release v1.2.0 adds a command-line tool with "file watcher and upload logic" and a background service, so a laptop folder can feed the workstation too; the notes give only these titles, so test it first.

A phone with the web app. Scriberr has a built-in recorder ("Use the built-in audio recorder to capture thoughts on the fly") and "supports PWA (Progressive Web App) installation", which means it can be installed from the browser like an app. On a phone that reaches the workstation, at the office or through the company VPN, the recording goes straight into Scriberr. The workstation must be served over HTTPS for this: browsers allow microphone access "only in secure contexts (HTTPS)", as MDN's documentation states. Without a connection, staff record in the phone's own recorder app and upload later.

Plaud devices the company already owns. Plaud's help centre says "WAV export is available on Plaud App", and with Cloud Sync off its security page says recordings stay on-device and the cloud is used only on demand for AI processing. By our reading, an exported WAV file can therefore be transcribed on your workstation without Plaud's transcription. The pages do not describe the export path in detail, so confirm with Plaud and your data-protection officer before you rely on it.

You can order the workstation alone, the setup of Scriberr, the speech models and the 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.

Summaries for client meetings, site visits and phone notes

In Scriberr the user picks a template prompt for each summary; what a company needs are templates for its own types of conversation, for example:

• client meeting: participants, the client's request, what was agreed, open points with owner and date;
• site visit: location, what was found, measurements and photo references mentioned, next steps;
• phone note for the CRM: contact, request, promised action, follow-up date.

The instruction tells the model to use only what was said and to mark unclear passages. Speaker labels read speaker 1 and speaker 2; a person adds names before the summary goes to a client or into the CRM.

Length is no problem. At an example 8,000 spoken words per hour and 1.76 tokens per German word, the figure cited in our guide to tokens per word in European languages, an hour of conversation comes to about 14,100 tokens, the units in which models count text. A session of 32,768 tokens for gpt-oss-20b, an open model under the Apache 2.0 licence, holds that transcript with template and summary. Parakeet TDT 0.6B v3 takes up to 3 hours of audio in one pass with local attention, according to its model card, in 25 European languages including German. Scriberr's README names Parakeet and Canary without a version or a language list, so for German recordings plan with Whisper and check which Parakeet version the installed app loads. Check the Canary model's licence as well: NVIDIA's first canary-1b is under CC BY-NC 4.0, which excludes commercial use, while canary-1b-flash is under CC BY 4.0. Scriberr's README also names a managed WhisperX environment for its models, and WhisperX's default alignment models for German, French, Spanish and Italian are under CC BY-NC 4.0, a non-commercial licence, so in a company set-up we check which alignment model the app loads and use one whose licence allows commercial use.

Recording conditions matter more than model size. Parakeet's model card gives 1.93 per cent word errors on read speech (LibriSpeech test-clean) against 11.31 per cent on meeting recordings (AMI). A recorder close to the speakers, or a clip-on microphone for the phone, does more than a larger model. Client and product names can be steered with a short vocabulary prompt, which OpenAI's guide to its speech service recommends for "names, acronyms, formatting, or recording-specific vocabulary".

Consent and the GDPR when recording client conversations

A recording of a conversation is personal data of everyone who speaks, and participants have to be informed and a legal basis has to exist before you record. In practice, say at the start that you would like to record, record only if everyone agrees, and never record a conversation secretly. Our meeting-minutes guide goes through informing participants and deleting recordings. National law also governs whether a conversation may be recorded at all; your legal counsel confirms what applies.

With a cloud recorder, the provider processes the recordings on your behalf as a processor under Article 28 GDPR, and the contract and the subprocessors belong in your review, as our guide to GDPR, cloud hosting and subprocessors explains. With the local chain, no outside provider transcribes or summarises. Your own duties stay: Article 32(1) names "the pseudonymisation and encryption of personal data" among the security measures, which for this chain means encrypted phones and recorders, deletion from the device after upload, an encrypted workstation and fixed deletion periods. Lawyers, tax advisers and auditors also have professional secrecy rules, set nationally and by their chambers, which your data-protection officer and your chamber confirm; our guide for law firms and tax advisers covers their other document work.

Which workstation for how many people

Transcription takes little GPU time. On one NVIDIA A100 in the Open ASR Leaderboard paper (arXiv 2510.06961v3), Whisper large-v3 transcribed 145.5 seconds of audio per second of computing and Parakeet TDT 0.6B v3 about 3,333. In an example of 20 people who record two hours a day each, the 40 hours of audio would take that card about 17 minutes with Whisper large-v3 and under a minute with Parakeet, by our arithmetic. No such figure is published for RTX PRO cards or for speaker labels, so a test with your recordings confirms the size.

Usable memory by our users-per-card rule (90 per cent of driver-visible memory, less 3 GiB): 18.5 GiB on the RTX PRO 4000, 25.7 GiB on the RTX PRO 4500, 83.0 GiB on a 96 GB card; Whisper turbo about 6 GB (5.6 GiB) and large about 10 GB (9.3 GiB) from OpenAI's repository; gpt-oss-20b 12.8 GiB with 0.75 GiB per 32K session, gpt-oss-120b 60.8 GiB with 1.125 GiB per 32K session (16-bit cache). Our estimates.

Entry. The NVIDIA RTX PRO 4000 Blackwell has 24 GB, draws 145 W and takes one slot. By our rule it leaves 18.5 GiB; Whisper turbo and gpt-oss-20b together take 18.4, which leaves no working room for both at once. The card suits transcription and speaker labels, with the summary as a second step once the speech model has freed its memory; the installation test shows how the app handles this. Parakeet, 0.6 billion parameters or about 1.1 GiB in 16-bit by our arithmetic, may fit next to gpt-oss-20b where the app's Parakeet version covers your languages; its card gives no GPU memory figure. Our page to buy an RTX PRO 4000 workstation lists the configuration.

Both models loaded. The RTX PRO 4500 Blackwell with 32 GB and 200 W leaves 25.7 GiB by our rule; after Whisper turbo and gpt-oss-20b, 7.3 GiB remain, room for 9 summaries of 32,768 tokens at the same moment. Both workstations run from an ordinary office socket.

Larger model and server. One RTX PRO 6000 Blackwell with 96 GB, as Workstation Edition or Max-Q, carries gpt-oss-120b next to Whisper large-v3, at about 1.1 kW at the wall with the Workstation Edition by our estimate. For a company-wide service, the entry server on our AI servers page, two RTX PRO 6000 Server Edition cards at about 1.8 kW by our estimate, runs the speech models next to the company assistant.

Send us the number of people who record, the hours per week and the languages, and we suggest a configuration.

Scriberr and Ollama in operation

1. Version and status. Scriberr's README has an undated status note, "Development is paused for now, but definitely not abandoned", next to a line announcing a restart; the install examples use v1.2.0. Pin that version and keep image and model files on your own storage. If development does not resume, the chain can be rebuilt from a speech server with an OpenAI-compatible endpoint and a workflow tool such as n8n with a local LLM.
2. Blackwell image. Scriberr's README says for RTX 50-series cards "You must use the Blackwell-specific image"; the RTX PRO Blackwell cards are Blackwell GPUs with compute capability 12.0, so we plan with that image, as our Otter-alternative comparison explains.
3. HTTPS, VPN and login. Serve Scriberr over HTTPS so phones can record, and reach it from outside only through the company VPN. The README does not describe user accounts or roles, so put it behind your company login on a reverse proxy.
4. Ports. Ollama binds to 127.0.0.1:11434 by default, but its container image sets OLLAMA_HOST=0.0.0.0; firewall the port so only Scriberr reaches it.
5. Speaker labels. Scriberr downloads PyAnnote models at first start; the gated pyannote community-1 pipeline needs its conditions accepted and an access token once, then runs offline, as our guide to speaker diarisation on-premise explains.
6. Hand-over. Release v1.2.0 integrates "webhook notifications into transcription pipeline", which can start a workflow that files the summary in your CRM or mail system.

FAQ