noScribe is the offline transcription app to choose when a recording must stay on your computer. It scores 4.0/5. Its README says it “runs completely locally”, and the pages we read describe no online features at all. The installer includes the AI models, so they don’t come as a separate download. It also labels speakers and has an editor for checking the text against the audio.
Buzz and Vibe transcribe locally too, but each has an optional feature that sends data to an online service: Buzz can use OpenAI’s API for speech recognition, and Vibe can summarize transcripts with the Claude API. Pick Buzz for the widest GPU support in a desktop app, Vibe for the most export formats, and whisper.cpp if you work from the command line.
Our picks
- Best overall noScribe 4.0 /5
- Best GPU support in an app Buzz 3.7 /5
- Most export formats Vibe 3.6 /5
- Best for the command line whisper.cpp 3.6 /5
Scores
1–5 per criterion · total is the weighted average · select a score to see its source| # | Product | What leaves your computer ×3 | Speaker labels and review ×3 | Export formats ×2 | GPU support and platforms ×2 | Total |
|---|---|---|---|---|---|---|
| 1 | noScribe | 5 | 5 | 3 | 2 | 4.0 |
| 2 | Buzz | 3 | 4 | 3 | 5 | 3.7 |
| 3 | Vibe | 3 | 3 | 5 | 4 | 3.6 |
| 3 | whisper.cpp | 4 | 2 | 4 | 5 | 3.6 |
How each criterion is scored
What leaves your computer ×3
Whether the project documents that transcription runs on your own computer, whether it has optional features that send audio or text to an online service, and whether it documents running with no internet at all.
- 5 As 4, plus the models ship with the download or the project documents a full no-internet setup
- 4 Transcribes locally, with no online features documented
- 3 Transcribes locally, with optional features that send audio or transcripts to an online service when you turn them on
- 2 Transcribes locally, but the documentation says data goes online by default
- 1 Audio is sent to an online service by default
Speaker labels and review ×3
Whether the transcript marks who is speaking (speaker identification, also called diarization), and whether the app has a viewer or editor for checking the text against the audio.
- 5 Speaker identification and an editor for correcting the text against the audio
- 4 Speaker identification and a transcript viewer with audio playback
- 3 Speaker identification or a transcript viewer, but not both
- 2 Experimental speaker-change markers only
- 1 Neither speaker labels nor a transcript viewer
Export formats ×2
How many file formats a transcript can be saved in, including subtitle formats (SRT, WebVTT) and formats a word processor opens (DOCX, HTML, PDF).
- 5 Six or more formats, including SRT, WebVTT and a word-processor format (DOCX, HTML or PDF)
- 4 Four or more formats, including SRT and WebVTT
- 3 Three formats, including at least one subtitle format
- 2 Two formats
- 1 Plain text only
GPU support and platforms ×2
Which graphics processors (GPUs) can speed up transcription (NVIDIA, Apple Silicon, AMD, Intel), on which operating systems, and whether it also runs on the processor (CPU) alone.
- 5 As 4, on Windows, macOS and Linux, with a documented processor-only mode
- 4 NVIDIA, Apple Silicon, and AMD or Intel GPUs (for example through Vulkan)
- 3 NVIDIA and Apple Silicon GPUs
- 2 One GPU family, such as NVIDIA only
- 1 Processor (CPU) only
noScribe was written for transcribing interviews in social research and journalism, and its privacy claim is the plainest of the four: “No cloud, no worries”. The documentation backs the claim with a detail that matters. Because the AI models are part of the download, the installer is about 3 GB, and the models don’t have to be fetched from anywhere later. The pages we read list no optional online features to audit.
The trade-off is speed. The project estimates 1 to 3 hours of processing per hour of interview, and an NVIDIA card with CUDA is the only documented way to speed that up. Output is an HTML file that opens in Microsoft Word, plain text, or WebVTT subtitles, and it can mark pauses and overlapping speech.
- Runs completely locally, and the models ship with the installer of about 3 GB
- Speaker recognition and support for around 60 languages
- An editor that aligns audio and text for corrections
- Free and open source under the GPL-3.0 license
- The only GPU speed-up documented is for NVIDIA cards (CUDA)
- One hour of interview takes about 1–3 hours to transcribe
- Three export formats, and the only subtitle format is WebVTT
Skip it if: Anyone who needs SRT subtitles, or fast results without an NVIDIA graphics card.
Buzz transcribes files and live microphone audio on Mac, Windows and Linux, where it comes as a Flatpak, Snap or AppImage. It documents the broadest hardware support of the three desktop apps: CUDA on NVIDIA cards, Apple Silicon, Vulkan “on most GPUs, including integrated GPUs”, and a Whisper.cpp mode that runs on the processor alone.
Local is the default, not the only option. Its FAQ says Buzz “also supports using OpenAI API to do speech recognition on a remote server” once you enter an API key, and then your audio goes to OpenAI. For a machine that never goes online, the FAQ describes downloading models on another computer and copying them over.
- GPU support for NVIDIA (CUDA), Apple Silicon and most other GPUs through Vulkan
- Speaker identification and a transcript viewer with search and playback controls
- Documented setup for a computer with no internet connection
- An optional OpenAI API mode does speech recognition on a remote server
- The Windows installer is not signed
- Current versions need an Apple Silicon Mac
Skip it if: Intel Mac owners, since Buzz now requires Apple silicon.
Vibe’s README promises “fully offline transcription, no data ever leaves your device”, and for transcription itself nothing in it contradicts that. The same feature list adds two exceptions you opt into: summaries through the Claude API, an online service, and transcribing audio from sites such as YouTube, which needs a connection. For summaries that stay local, it also supports Ollama.
Its strength is output. It saves SRT, VTT, TXT, HTML, PDF, JSON and DOCX, labels speakers, and can transcribe a batch of files. It ties whisper.cpp at 3.6/5 because the README lists a realtime preview but no viewer or editor for reviewing the result.
- Seven export formats, including SRT, VTT, DOCX and PDF
- Speaker diarization and batch transcription of several files
- GPU support for NVIDIA, AMD and Intel through Vulkan, and CoreML on macOS
- Optional summaries use the Claude API, so the transcript leaves your device for that step
- Lists automatic updates without saying what the update check sends
Skip it if: Anyone who needs an editor for correcting the transcript against the audio, which Vibe's README doesn't list.
whisper.cpp is the engine underneath this category. Vibe credits it, and Buzz offers it as one of its model types. On its own it is a library plus a command-line tool that you build with CMake, and it runs on macOS, Linux and Windows with the widest hardware list here. Nothing in the README describes an online feature; the models are a separate download.
It shares third place with Vibe at 3.6/5. Its hardware support earns the top score, but it is a tool for people who write scripts: speaker segmentation through tinydiarize is marked experimental, and there is no viewer. The command-line tool writes text, VTT, SRT, LRC, CSV and JSON files.
- Plain C/C++ with no dependencies, and no online features documented
- Six output formats from the command-line tool, including SRT and VTT
- Supports NVIDIA, AMD (ROCm), Vulkan, Apple Metal and Core ML, or the processor alone
- Speaker labels are experimental and only mark speaker changes
- The command-line example reads only 16-bit WAV files, so other audio must be converted first
- No graphical app; you build the command-line tool yourself
Skip it if: Anyone who wants a window, buttons and an editor.
Questions
Do any of these send my audio online?
Not by default, according to their documentation. Buzz sends audio to OpenAI only if you turn on its OpenAI API mode, and Vibe sends transcripts to the Claude API only if you use its summaries. The noScribe and whisper.cpp pages we read describe no online features. Buzz and whisper.cpp download their models separately; noScribe includes them in the installer.
Which ones label who is speaking?
noScribe, Buzz and Vibe all identify speakers. whisper.cpp has only experimental speaker-change markers through tinydiarize.
Which ones can make SRT subtitles?
Buzz, Vibe and whisper.cpp export SRT and VTT. noScribe exports WebVTT, plus HTML and plain text.
Method
Each tool was scored 1–5 against a published scale on four weighted criteria: what leaves your computer (×3), speaker labels and review (×3), export formats (×2) and GPU support and platforms (×2). The ranking is the weighted average. Every score cites quotes from the project’s own README, website or documentation, captured on 5 November 2026.
We did not install the tools, test their accuracy or watch their network traffic, so the privacy scores reflect what each project documents, not what we measured. How our rankings work.