Transcribe interviews locally with noScribe and a GPU, for research and journalism

Transcribing interviews on your own computer

Interviews can be transcribed locally with noScribe, a free app for Windows, macOS and Linux that turns a recording into a transcript with speaker labels and timestamps without sending anything to a cloud service. On an ordinary computer one hour of interview takes one to three hours, by the project's own figure. An NVIDIA graphics card shortens that wait, and a workstation with one NVIDIA RTX PRO 4000 Blackwell is the entry point for a researcher or a small team.

This guide is for market research agencies, social research institutes, editorial offices and consultancies that run user or expert interviews and have promised their interviewees confidentiality. It explains what noScribe does, how long it takes with and without a GPU, what data protection asks of interview recordings, how the transcript gets into MAXQDA or Word, and which workstation fits which volume. Interviews with retiring staff at the machine are covered in our guide to capturing expert knowledge before retirement; the speaker-labelling models themselves in our guide to speaker diarisation on-premise.

Some terms first. A transcript is the written text of the interview. Speaker labels, called diarisation by engineers, mark which passage belongs to which voice. A timestamp is the time in the recording at which a passage starts. CUDA is NVIDIA's software layer through which an app uses the graphics card (GPU). QDA software, such as MAXQDA or ATLAS.ti, is used to code and analyse qualitative data.

What noScribe does with an interview

The project's README describes noScribe as "An app to produce high quality transcripts of interviews for qualitative social research or journalistic use". It is published under GPL-3.0 and uses Whisper from OpenAI through faster-whisper for the speech recognition and pyannote for the speaker labels. Its site, noscribe.de, states: "No cloud, everything stays on your own computer." The installer is about 3 GB because it includes the AI models, so after installation the app needs no internet connection for transcription.

noScribe README on GitHub, read 30 October 2026.

After the run, the editor opens the transcript next to the recording. Ctrl and the space bar play the audio at the cursor position, at adjustable speed, so names, numbers and unclear passages can be checked where they occur. A "Queue" tab lists all jobs with their state. Version 0.7 added processing of several files at once and a command line interface. The latest release on GitHub as of October 2026, 0.7.2, is a maintenance release that adds a "True headless mode for servers", while the README still gives 0.7 as the current version number.

The output is an HTML file by default, which "can be opened in every common word editor", plus plain text or WebVTT, a subtitle format the project added for import into EXMARaLDA. On languages the README is plain: Whisper "understands around 60 different languages, but the quality of the transcription varies widely between them", with Spanish, Italian, English, Portuguese and German the best supported. Our comparison of German speech recognition, accuracy and dialects covers regional speech.

One warning comes from the project itself. Its site states: "Someone is trying to sell paid transcription services via the domain noscribe(dot)ai. I have nothing to do with that." The official site is noscribe.de, and the code is on GitHub under kaixxx/noScribe.

How long transcription takes on a CPU and with a GPU

noScribe's site gives the planning figure for an ordinary computer: "One hour of interview takes about 1-3 hours of transcription time, on older machines even longer", and suggests queueing jobs "for example overnight". It adds that "An NVIDIA graphics card (CUDA) can significantly speed up transcription." The Windows build for this, according to the README, needs "NVIDIA graphics cards with at least 6 GB of VRAM", NVIDIA drivers "on version 570.65 or higher" and an installed CUDA toolkit.

noScribe publishes no timing on any GPU. Published figures exist for the engines underneath. The Open ASR Leaderboard paper measured Whisper large-v3 on one NVIDIA A100-SXM4-80GB at an RTFx of 145.5, meaning 145.5 seconds of audio per second of computing, with batches of 64; by our arithmetic, recognising a 60-minute interview would take about 25 seconds on that card. WhisperX, another tool built on faster-whisper, states "70x realtime transcription using whisper large-v2" in batched mode without naming the GPU, and "<8GB gpu memory for large-v2 with beam_size=5". These figures come from batched set-ups on other hardware and leave out speaker detection, so they do not tell you noScribe's time on a workstation. What does decide it: the "precise" or "fast" setting, speaker detection on or off, the number of speakers, the audio quality and the card. Time one of your typical recordings on the delivered machine, and plan the queue from that.

The project names no tested list of graphics cards. On a new Blackwell card, install a driver at or above the version noScribe requires, and a first test transcription shows whether the app uses the GPU.

You can order the workstation alone, the setup of noScribe 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.

Interview data, consent and the GDPR

Research and journalistic interviews often touch the categories in Article 9(1) of the GDPR, which names data "revealing racial or ethnic origin, political opinions, religious or philosophical beliefs, or trade union membership" and "data concerning health", among others. A voice recording with a name attached holds personal data in any case, and so does the transcript until it is pseudonymised.

When a cloud service transcribes the recordings, it processes them on your behalf, and Article 28(3) requires a contract that "sets out the subject-matter and duration of the processing, the nature and purpose of the processing, the type of personal data and categories of data subjects". One university page shows both routes side by side. Universität Hamburg's Faculty of Education offers the automatic transcription service f4x to its staff, which its page describes as an offer of Dr. Dresing & Pehl GmbH with servers in Germany and encryption; the same page notes that all rights needed for processing by third parties in the cloud must be in place, and points to the free noScribe for local transcription. Which route a project may take depends on what interviewees were told and what they agreed to.

Three steps help on either route. Check what the consent form and the data-management plan say about who transcribes and where; local transcription adds no processor. Store the audio on an encrypted drive and delete it when the consent or the project plan says so; Article 32(1) lists "(a) the pseudonymisation and encryption of personal data" among the security measures. And pseudonymise the transcript early. Article 4(5) defines pseudonymisation as processing so that the data "can no longer be attributed to a specific data subject without the use of additional information", provided that information "is kept separately". In practice names, employers and places are replaced by codes, and the key list sits in a separate place. A local language model can suggest the replacements, and a person checks them, as our guide to PII redaction before an LLM describes.

Rules on research exemptions and on the protection of journalistic sources are partly set by national law. Your data-protection officer, ethics committee or legal counsel confirms what applies to your project. This section is general information as of October 2026, not legal advice.

From transcript to MAXQDA, ATLAS.ti or Word

noScribe's README says the HTML file opens in common word processors, Word and LibreOffice included, and in QDA packages such as MAXQDA, ATLAS.ti and QualCoder. For MAXQDA the route with links to the audio runs through a DOCX or RTF file. MAXQDA's online help, read on 30 October 2026, states that "MAXQDA supports the formats of nearly all transcription programs", lists the timestamp formats it recognises, among them [hh:mm:ss] and the #hh:mm:ss-x# format of f4x, and converts them into its own timestamps. It advises keeping "the transcripts and the media files in the same folder, using the same names", and asks for the media file during import.

By our reading, noScribe's [hh:mm:ss] timestamps match one of the listed formats. The steps are: switch timestamps on in noScribe, open the HTML file in Word or LibreOffice, save it as DOCX next to the audio file with the same name, and import it in MAXQDA as a transcript with timestamps. Test this with one interview before a project starts. The VTT export serves EXMARaLDA and other tools that read subtitle files; MAXQDA's help also lists VTT subtitle files, such as those from Teams, Zoom or WebEx, among the formats it imports, so noScribe's VTT file is a second route to test.

Transcription rules come before the software. A project decides whether it needs a verbatim transcript with fillers, pauses and broken-off words or a cleaned one. noScribe can mark pauses as "(..)" and attempts fillers, which its README calls "more of a 'recommendation'" to the model, so a person still corrects the transcript against the audio in the editor. Universität Hamburg's page describes the same step for f4x: users review the raw transcript and add what their transcription guidelines require.

Which workstation for interview transcription

noScribe is a desktop app, so the first tier is a workstation for the person or team that transcribes. A language model joins when the team wants summaries, a first set of codes or answers across many interviews; an hour of interview at an example 8,000 words comes to about 14,100 tokens in German, as our guide to meeting minutes with a local AI works out, so one session of 32,768 tokens holds a whole interview with the instruction and the answer.

GPU memory from NVIDIA's datasheets; noScribe minimum from its README; Whisper memory from OpenAI's repository (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 40.0 GiB usable on the 48 GB RTX PRO 5000 and 83.0 GiB on a 96 GB card; 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, as in our model guides. Our estimates, before the speaker model, whose memory its card does not state.

Entry. One NVIDIA RTX PRO 4000 Blackwell has 24 GB of GDDR7 at 145 W in a single slot, four times the minimum noScribe asks for and more than the about 10 GB OpenAI gives for Whisper large. It suits one researcher, or a small team that shares the machine and lets the queue run overnight. Our page to buy an RTX PRO 4000 workstation has the card's data.

Entry with summaries. The NVIDIA RTX PRO 5000 Blackwell with 48 GB at 300 W adds a local language model. By our rule it leaves 40.0 GiB; gpt-oss-20b takes 12.8 and Whisper about 9.3, and the remaining 17.9 GiB hold 23 sessions of 32,768 tokens, so several team members can ask for summaries or first codes at the same time. The model's output is a draft for the researcher, never the analysis itself.

Mid. A workstation with one NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB, 600 W) runs the larger gpt-oss-120b next to the transcription and draws about 1.1 kW at the wall by our estimate, from an ordinary office circuit. A second card of the same edition doubles the room for parallel work; two are the practical maximum for that edition.

Large. When many projects transcribe at once and staff should upload recordings through a browser, the work moves to a shared server. Our comparison of a self-hosted Otter alternative covers the web apps for that, and our AI servers page the hardware.

Send us the number of interviews a month and their languages, and we suggest the best solution for your tasks and budget.

Setting up noScribe on a GPU workstation

1. CUDA build. Windows: the CUDA installer needs at least 6 GB of VRAM, NVIDIA driver 570.65 or later and an installed CUDA toolkit, followed by a reboot (README). RTX PRO Blackwell cards have compute capability 12.0; check with a first transcription that the GPU is used.
2. Versions. README version 0.7; latest GitHub release 0.7.2, a maintenance release with headless mode. The command line takes noScribe.exe --help and --no-gui for scripts. Release 0.6 added custom Whisper models.
3. Offline operation. The models ship in the installer of about 3 GB, so transcription needs no outside connection; block outbound traffic from the workstation if your rules require it.
4. Shared server. WhisperX 3.8.6 (25 May 2026, BSD-2-Clause) adds word timestamps and pyannote diarisation, which needs a Hugging Face token and acceptance of the model's conditions. Its default alignment models for German, French, Spanish and Italian are torchaudio's VoxPopuli models under CC BY-NC 4.0, a non-commercial licence, so a company server needs an alignment model whose licence allows commercial use. Scriberr (MIT, v1.2.0) states "Development is paused for now, but definitely not abandoned", announces a restart in the same README and needs its Blackwell-specific image on these cards.
5. Storage. Encrypt the data drive, include it in the backup plan and set deletion dates per project.
6. One card, two programs. If the language model runs in vLLM next to noScribe, lower --gpu-memory-utilization from its default of 0.92, because vLLM reserves that share of the card's memory when it starts and leaves too little for Whisper.

FAQ