Local AI dictation as a Dragon alternative for offices and law firms, without the cloud

Local AI dictation instead of Dragon

Local AI dictation works without any cloud service: an open dictation app records while you hold a key, a speech recognition model turns the audio into text, and the text appears in Word, the case-management system or any other field. For a few people the model can run on each PC. For an office or law firm with many users, thin clients or terminal servers, one workstation with an NVIDIA RTX PRO 4000 Blackwell card runs the model for everyone, and the audio stays inside the company network.

This guide is for the office manager, the partner in a law firm or the head of a secretariat who dictates letters, briefs and reports with Dragon Professional Anywhere or another cloud app today and wants the audio to stay in the building, or who looks for an open alternative to Dragon. Our guide to the speech-to-text server for Whisper and Parakeet sizes servers for recorded audio and call centres; here the subject is dictation into a text field.

Four terms come up throughout. A speech recognition model turns speech into text; Whisper from OpenAI and Parakeet from NVIDIA are the two used here. A GPU, or graphics card, runs such a model quickly. An endpoint is an address on the company network that receives audio and returns text. A language model is a second AI that rewrites text, for example a dictated draft into a formatted letter.

What Dragon Professional Anywhere does today

Nuance's product page, read on 30 October 2026, describes Dragon Professional Anywhere as "cloud-hosted speech recognition for the enterprise and public sector", with data "encrypted in transit and at rest". Administrators "manage or share customizations, including custom words, commands and auto-texts, across multiple users", and the page names "support for thin client hardware, server virtualization, as well as Citrix environments". The release notes for version 2026.1 list Windows 10 and 11 and Windows Server 2016 to 2025 as supported systems. The product page names no region where the speech is processed.

With this product the speech is recognised in Nuance's cloud, so the audio of every dictation leaves the company. For many offices that is the reason to look for a local alternative. Nuance's product page and release notes give no end-of-sale date for the product. They do set a life cycle per release: end of life is "now 24 months after the initial release date", and the table in the 2026.1 notes lists 13 February 2027 for release 2025.1. Your Dragon reseller can confirm in writing which dates apply to the version you run.

Two ways to dictate locally: on each PC or through one workstation

Handy README and release notes on GitHub; OpenWhispr README and models page; vLLM speech-to-text documentation of 19 October 2026; all read 30 October 2026. The comparison is ours.

On each PC. Handy is "a free, open source, and extensible speech-to-text application that works completely offline": you press a shortcut, speak, and "have your words appear in any text field". Its README states "The process is entirely local". It offers Whisper models from Small to Large, which use the graphics card "when available", and Parakeet V3, a "CPU-optimized model" with automatic language detection; Silero VAD filters out silence first. Handy runs on Windows, macOS and Linux under the MIT licence; its latest release as of October 2026 is 0.9.6 of 24 August 2026. For three or four people with current PCs this needs no new hardware and is the right place to start.

Through one workstation. The second way puts the model on one machine in the office. A transcription service runs there and offers an endpoint compatible with OpenAI's transcription interface; vLLM's documentation describes it as /v1/audio/transcriptions, "compatible with OpenAI's Transcriptions API". On the PCs, a dictation app sends each recording to that address instead of recognising it itself. OpenWhispr, under the MIT licence for Windows, macOS and Linux, supports this: its models page lists a custom endpoint, "Any OpenAI-compatible API", as a transcription provider for "self-hosted Whisper servers" and "any service with an OpenAI-compatible /audio/transcriptions endpoint", and its README states "No data collection, no telemetry". The same app also offers cloud transcription, cloud AI models and an optional OpenWhispr Cloud account; in the office set-up these stay off and the app uses only the workstation's endpoint.

The workstation pays off on thin clients, terminal servers and PCs without a capable graphics card, where the model would otherwise run on shared or slow processors; with the endpoint, the session only passes on the audio. Every user also gets the large Whisper model, word lists and updates live in one place, and one language model formats letters for the whole office. In return, audio crosses the internal network, the workstation has to run when people dictate, and someone owns its updates.

Test virtual desktops before a rollout. Dragon's release notes list the Citrix and virtual-desktop client versions it supports; with an open dictation app, a short pilot shows whether the microphone reaches the session and whether the text lands in the right window.

You can order the workstation alone, the setup of the speech model, the transcription service and OpenWhispr for the PCs 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.

Whisper or Parakeet for dictating letters and briefs

Parameters and memory from OpenAI's Whisper repository and NVIDIA's model card; German word error rate (WER) from the Open ASR Leaderboard paper (arXiv 2510.06961v3), measured on an NVIDIA A100-SXM4-80GB; all read 30 October 2026.

OpenAI's repository releases Whisper's code and weights under the MIT licence and describes turbo as offering "faster transcription speed with a minimal degradation in accuracy". NVIDIA publishes Parakeet TDT 0.6B v3 under CC BY 4.0, "ready for commercial/non-commercial use"; it transcribes 25 European languages, among them German, French, Italian, Spanish, Dutch, Polish and Czech, with punctuation and capital letters. Both models cover German, English and many other European languages.

The word error rate (WER) counts the share of words a model gets wrong. On public German test sets both models stay around 5 per cent, about one word in twenty. Dictation is the friendly case, because one person speaks close to the microphone. Parakeet's model card shows the difference: 1.93 per cent on LibriSpeech test-clean, which is read speech, against 11.31 per cent on AMI meeting recordings. Its noise test averages 6.34 per cent on clean audio and 11.66 per cent when the noise is as loud as the speech. A headset or a dictation microphone in a quiet room does more for accuracy than a larger model.

Whisper's model card adds two warnings. "Our models perform unevenly across languages", and also on "different accents and dialects", and "the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)". Every dictated letter is therefore read before it goes out, as it was with Dragon. Our comparison of German speech recognition, accuracy and dialects goes through the models for regional speech.

Names, legal terms and formatted letters

Dragon users rely on custom words, voice commands and auto-texts. The open tools cover them differently.

Custom words. Handy's release notes for 0.9.6 include a fix to its custom-words feature, and OpenWhispr's comparison page lists "Custom vocabulary support". On the workstation, vLLM's transcription interface accepts an optional prompt, "Optional text to guide the transcription style"; a short list of client names and legal terms can be passed there. Test with twenty of your own dictations whether it improves the spelling of the names that matter to you.

Voice commands. Nuance's page for Dragon Professional v16 says "Use custom voice commands to insert standard boilerplate text or signatures into documents". The READMEs of Handy and OpenWhispr describe dictation into the active text field; we found no documented equivalent of Dragon's command set in either. Offices that control their PCs by voice should keep that need in the pilot.

Auto-texts and letter formats. Here a language model takes over. The raw dictation goes to the model with an instruction such as "write this as a letter to the court, keep every fact, mark anything unclear". OpenWhispr's models page lists local models for this "AI text cleanup", among them GPT-OSS 20B, run "via llama.cpp with GGUF quantization" on the PC. On the office workstation the same model, gpt-oss-20b under Apache 2.0, serves everyone: its weights take 12.8 GiB, as our guide to 30B-class models on 24 GB cards records. A language model can change meaning, so the person who signs the letter checks it against the dictation.

Dictation, client confidentiality and the GDPR

Dictated letters in a law firm, a tax practice or an insurance office carry personal data of clients, opposing parties and witnesses, sometimes of a kind Article 9(1) of the GDPR treats separately, such as "data concerning health". When a cloud service recognises the speech, it processes that data on your behalf, and Article 28(3) applies: "Processing by a processor shall be governed by a contract or other legal act under Union or Member State law", which sets out the subject matter, duration, nature and purpose of the processing.

With local recognition the audio goes to no outside provider for that step. The duties of the firm stay. Article 32(1) asks for measures "to ensure a level of security appropriate to the risk", including "(a) the pseudonymisation and encryption of personal data". For a dictation workstation that means an encrypted connection from the PCs, named administrator accounts, and a service that keeps no audio after returning the text, with logs that record times and users without content. Our guide to local AI and the GDPR covers what own hosting changes, and our check of telemetry in self-hosted AI tools shows how to verify that no component calls outside services.

Rules on professional secrecy for lawyers, notaries, tax advisers and doctors are set nationally and by the professional chambers. Medical practices find doctor's letters and the medical-device question in our guide to medical dictation with local AI. Your data-protection officer and, where relevant, your chamber confirm what applies to dictation. This section is general information as of October 2026, not legal advice.

Hardware for local dictation, from PC to server

Dictation is a light load for a GPU. The size of the machine follows from how many people finish a dictation at the same moment and whether a language model also formats the text. No published benchmark gives dictation users per RTX PRO card; what exists is the A100 figure from the Open ASR Leaderboard paper: Whisper large-v3 transcribed 145.5 seconds of audio per second of computing, Parakeet TDT 0.6B v3 about 3,333. By our arithmetic on that card, one minute of dictation needs about 0.4 seconds of GPU time with Whisper and well under a tenth of a second with Parakeet. The paper processed up to 64 recordings at once, so a single dictation on its own waits longer than this share of GPU time. A pilot with your staff, logging the wait after each dictation, sets the final size.

GPU memory from NVIDIA's datasheets; Whisper memory from OpenAI's repository (6 GB = 5.6 GiB, 10 GB = 9.3 GiB); sessions by our users-per-card rule (90 per cent of driver-visible memory, less 3 GiB per card, less the weights), with 25.7 GiB usable on the RTX PRO 4500 and 83.0 GiB on a 96 GB card; gpt-oss-20b 12.8 GiB with 0.1875 GiB per 8K session, gpt-oss-120b 60.8 GiB with 1.125 GiB per 32K session (16-bit cache), as in our model guides. Our estimates, not measurements.

Entry. One NVIDIA RTX PRO 4000 Blackwell with 24 GB of GDDR7 and 145 W, in one slot, is enough for Whisper as the office dictation service. Our page to buy an RTX PRO 4000 workstation lists its data. When the language model for letters joins, the 32 GB RTX PRO 4500 Blackwell at 200 W is the step up: by our rule it leaves 25.7 GiB, of which gpt-oss-20b takes 12.8 and Whisper turbo about 5.6, and the remaining 7.3 GiB holds 38 formatting requests of 8,192 tokens at once, ample for letters of a page or two. Both workstations run on an ordinary office socket.

Mid. A workstation with one NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB, 600 W) carries the larger gpt-oss-120b next to Whisper. The firm then has one machine for dictation, letter formats and an internal assistant; by our estimate it draws about 1.1 kW at the wall. Our guide for law firms and tax advisers covers the document work such a model does.

Large. For several sites or a terminal-server farm with hundreds of users, the service moves to a rack server with L4 cards or RTX PRO 6000 Server Edition cards, sized on our AI servers page. It needs a server room, as our guide to a GPU server in an office explains.

Tell us how many people dictate and on which systems, and we suggest the best solution for your tasks and budget.

Endpoint, access and clients for the IT team

1. Endpoint. vLLM serves /v1/audio/transcriptions after pip install vllm[audio]; uploads are limited to 25 MB by default through VLLM_MAX_AUDIO_CLIP_FILESIZE_MB, and response formats include json, text and verbose_json. A Realtime API streams PCM16 audio at 16 kHz mono over WebSocket at /v1/realtime (vLLM documentation of 19 October 2026).
2. Alternative server. Speaches, under MIT, offers an OpenAI-compatible server on faster-whisper with "GPU and CPU support". Its latest release on GitHub is 0.9.0-rc.3, a release candidate; pin the version you deploy.
3. Access. vLLM's --api-key protects only the paths under /v1, /v2, /inference and /cohere, so the transcription path is covered, while /metrics, /health and /invocations answer without the key, and /invocations gives the same inference as /v1 (vLLM 0.32.0 security guide). Keep the service on the internal network behind an LLM gateway with keys and a query log, with TLS between the PCs and the workstation.
4. Clients. OpenWhispr is pointed at the endpoint as an OpenAI-compatible transcription provider; Handy's README names no server option, so it suits local use only. Handy's README notes that models can be installed manually for offline or restricted networks.
5. Blackwell builds. RTX PRO Blackwell cards have compute capability 12.0, and container images must be built for it; Scriberr's README, for example, says "You must use the Blackwell-specific image". Our comparison of a self-hosted Otter alternative covers transcription apps with uploads and speaker labels.

FAQ