Interview transcription you can correct and export: noScribe, ElevenLabs Scribe and three alternatives

noScribe transcribes interviews on your own computer for free, with speaker labels and a Word-ready file. ElevenLabs Scribe has the strongest public accuracy evidence, if you accept cloud processing.

PRACTICAL PICKSPreferredPickOur Preferred PickInterviews kept on your own computer$0

noScribe

Free, open-source transcription with speaker detection, optional timestamps and an editor that plays the audio from your cursor, all without uploading the recording.

Slow on ordinary computers: about 1–3 hours per interview hour. Windows or Apple silicon Mac; some Windows installs fail.

Why noScribe is our Pick
PRACTICAL PICKSRisingStarOur Rising StarAccuracy first, in the cloud$22/mo

ElevenLabs Scribe

A 2025 entrant whose Scribe v2 model had the lowest long-form error rate in an independent benchmark; its web app now uses the newer v3, not independently tested. The editor moves passages between speakers.

Cloud processing; switch off model training before your first upload. Retention of web-app uploads is not documented.

Why ElevenLabs Scribe earned Rising Star

Who is this article for? People with permission to transcribe recorded interviews, such as journalists, researchers and students, who need corrected, speaker-labelled text with timestamps. Not legal advice on consent or restricted recordings.

About 29 min read

What Preferred Pick, Rising Star and Ready to Rock mean

The quick list

Four one-hour interviews a month, transcribed and corrected. Compare the normal price, where the recording goes and how you fix speakers.

Normal prices for four one-hour interviews a month, checked November 12, 2026. Our arithmetic, before taxes and promotions. ElevenLabs hours are its Creator credits divided by its published transcription rate.

The comparison at a glance

Compare where the recording is processed, public accuracy evidence, how you fix speakers, what you can export and what four interviews a month cost.

Order follows our verdict for permissioned recorded interviews, not a measured accuracy ranking. All tools remain selectable.

Transcription tools for permissioned recorded interviews
FeaturenoScribeElevenLabs ScribeMacWhisperVibeBuzz
Best fitPreferred Pick: permissioned interviews kept on your own Windows PC or Apple silicon MacRising Star: accuracy first, if cloud processing is acceptableAlso consider: Mac users who want faster local processing and fuller speaker editingFree local transcription with many export formats; speaker correction not documentedFree local transcription with speaker identification; text and subtitle exports only
Where the audio is processedOn your computer; works without a network connection, per the developerElevenLabs serversOn your Mac by default; optional cloud providers receive audioOn your computer, per its README; model downloads and updates use the networkOn your computer, per its README
Four one-hour interviews a month$0Creator $22/month, or $220 billed yearly (about 6 hours of credits)Pro €65 once$0$0
Free routeEntirely free (GPL-3.0)About 30 minutes a month; no commercial licenceFree edition: smaller models; up to two speakers added by handEntirely free (MIT)Entirely free (MIT)
Public accuracy evidenceUses Whisper V3 Turbo: 11.0% long-form and 17.9% interview word error rate on the Open ASR Leaderboard (model, not this app)Web app now defaults to Scribe v3, not independently tested; its predecessor Scribe v2 had the lowest error rates on the same leaderboard, 7.3% long-form and 10.7% on interviewsLocal Parakeet v3 (15.6% on interviews) or Whisper Large v3 (18.9%); cloud Scribe optionalWhisper, Parakeet and Nemotron models; default not stated; Parakeet v3 scored 15.6% on interviews (model, not this app)Whisper family; version not stated; no benchmark of the app
Speaker correctionEdit labels as text or with find and replace; custom names since 0.8Rename speakers; move segments to another speaker, singly or in bulkPro: rename, merge and reassign with number keysNot documented in the pages we readNot documented in the pages we read
Document export and timestampsHTML that opens in Word, TXT; timestamp every 60 seconds or at speaker changesTXT, HTML, JSON; word-level timing in the editorDOCX, PDF, TXT, Markdown, HTML, CSV; timecodesSRT, VTT, TXT, HTML, PDF, JSON, DOCX; timestamp and speaker options not describedTXT, SRT, VTT
Training on your uploadsNone: nothing leaves your computerOn by default; switch off under Data use before uploadingNone locally; a cloud provider’s terms apply if you choose oneNone locally; the optional Claude summary sends the transcript to AnthropicNone locally
PlatformsWindows 10–11; macOS 14+ on Apple silicon; LinuxWeb browsermacOS 14+ (Apple silicon recommended); separate App Store editionmacOS, Windows and LinuxWindows (unsigned installer), macOS on Apple silicon, Linux
Main cautionSlow on ordinary computers; some Windows installs failTraining on by default; web-app retention undocumentedMac only; its 2024 privacy policy is older than its cloud featuresSpeaker correction and privacy policy not read; pull requests limited to a few collaboratorsNewest speaker feature; text and subtitle exports only; unsigned Windows installer

Permissioned recorded interviews. Prices checked November 12, 2026; other facts checked October 19, 2026 unless a source says otherwise; next review December 18, 2026, or sooner after a material change. Prices assume four one-hour interviews a month at normal rates, before taxes and promotions. This is public-source research: we uploaded no recordings, created no accounts and installed nothing. Word error rates describe speech models on public test sets, not your recordings. Check every quotation and speaker label against the audio.

All features and sources 20 criteria

Open the groups that matter to you. They compare the same selected tools. “Planned” means future work, not an available feature. Sources include the published conditions.

Cost, plans and platforms

Show 7 criteria
Cost, plans and platforms
FeaturenoScribeElevenLabs ScribeMacWhisperVibeBuzz
Best fitPreferred Pick: permissioned interviews kept on your own Windows PC or Apple silicon MacRising Star: accuracy first, if cloud processing is acceptableAlso consider: Mac users who want faster local processing and fuller speaker editingFree local transcription with many export formats; speaker correction not documentedFree local transcription with speaker identification; text and subtitle exports only
Four one-hour interviews a month$0Creator $22/month, or $220 billed yearly (about 6 hours of credits)Pro €65 once$0$0
Free routeEntirely free (GPL-3.0)About 30 minutes a month; no commercial licenceFree edition: smaller models; up to two speakers added by handEntirely free (MIT)Entirely free (MIT)
Paid plan limitsNo plans; queue several files, for example overnightCredits shared with other ElevenLabs tools; transcription uses 330 credits a minute; Creator has 121,000 a monthPersonal licence with a soft limit of three devicesNo plans; donations optionalNo plans
File limitsNone stated; long files take proportionally longer3 GB per fileNone stated; depends on your MacNone stated; batch transcription of several filesNone stated
LanguagesAbout 60; quality varies by language90+About 100 with Whisper; varies by modelAlmost every language, per its README; quality varies by modelWhisper languages; not listed in its README
PlatformsWindows 10–11; macOS 14+ on Apple silicon; LinuxWeb browsermacOS 14+ (Apple silicon recommended); separate App Store editionmacOS, Windows and LinuxWindows (unsigned installer), macOS on Apple silicon, Linux
Sources for cost, plans and platforms

Prices checked November 12, 2026: normal prices for four one-hour interviews a month, our arithmetic, before taxes and promotions. noScribe download and requirements; ElevenLabs pricing and speech-to-text guide; MacWhisper sales page, site and licensing; Vibe; Buzz. ElevenLabs hours are credits divided by its published rate.

Accuracy, speakers and correction

Show 5 criteria
Accuracy, speakers and correction
FeaturenoScribeElevenLabs ScribeMacWhisperVibeBuzz
Public accuracy evidenceUses Whisper V3 Turbo: 11.0% long-form and 17.9% interview word error rate on the Open ASR Leaderboard (model, not this app)Web app now defaults to Scribe v3, not independently tested; its predecessor Scribe v2 had the lowest error rates on the same leaderboard, 7.3% long-form and 10.7% on interviewsLocal Parakeet v3 (15.6% on interviews) or Whisper Large v3 (18.9%); cloud Scribe optionalWhisper, Parakeet and Nemotron models; default not stated; Parakeet v3 scored 15.6% on interviews (model, not this app)Whisper family; version not stated; no benchmark of the app
Speaker detectionpyannote; set the speaker count to avoid extra speakersUp to 32 speakersPro: supported local models on M-series Macs, or ElevenLabs and Deepgram in the cloudSpeaker diarization listed; its newest model handles up to 8 speakersSpeaker identification on finished transcripts; you can set the speaker count
Speaker correctionEdit labels as text or with find and replace; custom names since 0.8Rename speakers; move segments to another speaker, singly or in bulkPro: rename, merge and reassign with number keysNot documented in the pages we readNot documented in the pages we read
Editor and playbackSeparate editor; playback follows the cursor; seek bar since 0.8Text editor with segment timing and word re-alignmentBuilt-in editor with synced playbackLive preview; no transcript editor described in the pages we readTranscription viewer with search, playback controls and speed adjustment
Processing and setupAbout 1–3 hours per interview hour on an ordinary computer, per the developer; about 3 GB installerUpload; processed on ElevenLabs serversVendor claims up to about 30× real time with Whisper; faster with ParakeetRuns on your computer; graphics-card acceleration where available; speed not statedRuns on your computer; Nvidia, Apple silicon and Vulkan acceleration; speed not stated
Sources for accuracy, speakers and correction

Word error rates are the Open ASR Leaderboard long-form English results, updated September 30, 2026 (method paper); “interviews” is its CORAAL subset of recorded sociolinguistic interviews. They measure speech models, not these apps’ full workflows. Model use: noScribe models, ElevenLabs models, MacWhisper sales page, Vibe, Buzz. Editing: noScribe usage, ElevenLabs transcripts, MacWhisper speakers. Processing speeds are vendor or developer statements.

Exports

Show 2 criteria
Exports
FeaturenoScribeElevenLabs ScribeMacWhisperVibeBuzz
Document export and timestampsHTML that opens in Word, TXT; timestamp every 60 seconds or at speaker changesTXT, HTML, JSON; word-level timing in the editorDOCX, PDF, TXT, Markdown, HTML, CSV; timecodesSRT, VTT, TXT, HTML, PDF, JSON, DOCX; timestamp and speaker options not describedTXT, SRT, VTT
Subtitle filesVTTSRT, VTTSRT, VTTSRT, VTT, with a stable-timestamp modeSRT, VTT
Sources for exports

Documented formats, not export tests: noScribe usage; ElevenLabs transcripts; MacWhisper sales page; Vibe; Buzz.

Data handling and exit

Show 6 criteria
Data handling and exit
FeaturenoScribeElevenLabs ScribeMacWhisperVibeBuzz
Where the audio is processedOn your computer; works without a network connection, per the developerElevenLabs serversOn your Mac by default; optional cloud providers receive audioOn your computer, per its README; model downloads and updates use the networkOn your computer, per its README
Training on your uploadsNone: nothing leaves your computerOn by default; switch off under Data use before uploadingNone locally; a cloud provider’s terms apply if you choose oneNone locally; the optional Claude summary sends the transcript to AnthropicNone locally
Deleting your filesYour files and backupsTranscripts can be deleted; logs may keep related data; zero retention is Enterprise-only and never covers the web appYour files and backupsYour files and backupsYour files and backups
Legal and incident recordNone found; open installation bug reportsClass action filed May 2026 alleging voice-model training without consent (allegations); earlier voice-actor suit settled 2025None foundNot checked beyond its repository pageNot checked beyond its repository page
Who runs itKai Dröge and contributors; open sourceEleven Labs Inc.; ElevenLabs Poland sp. z o.o. for Europe and the UKGood Snooze (Jordi Bruin), NetherlandsThe thewh1teagle GitHub account and contributors; open sourceThe chidiwilliams GitHub account and contributors; open source
Main cautionSlow on ordinary computers; some Windows installs failTraining on by default; web-app retention undocumentedMac only; its 2024 privacy policy is older than its cloud featuresSpeaker correction and privacy policy not read; pull requests limited to a few collaboratorsNewest speaker feature; text and subtitle exports only; unsigned Windows installer
Sources for data handling and exit

What matters when choosing

noScribe is our Practical Pick for someone with permission to transcribe recorded interviews who wants a corrected, speaker-labelled transcript without sending the recording to a company. It is free, open-source software for Windows, Apple silicon Macs and Linux, written by a sociologist for research and journalistic interviews. It runs OpenAI’s Whisper V3 Turbo speech model and the pyannote speaker-detection model on your own computer, and its developer says it works with the network switched off. It saves an HTML transcript that opens in Word, with speaker labels and, if you choose, a timestamp every minute and at each change of speaker. Its editor plays the recording from wherever your cursor sits, so you can check a passage against the audio. The most important compromise is time: the developer estimates one to three hours of processing for each hour of interview on an ordinary computer, and the installer is about 3 GB. Correcting a speaker means editing the label text or using find and replace; there is no reassign button. Its lead is narrow. noScribe scored 3.95 out of 5 against 3.86 for MacWhisper, which runs only on Macs, and 3.73 for Vibe, whose scores rest on its project page alone; both gaps are smaller than one point on our most heavily weighted criterion. What noScribe is, usage guide, FAQ, models, download page

Accuracy is the job, and no public source has tested these tools on the same interviews. The best independent evidence measures the speech models underneath them. On the Open ASR Leaderboard’s long-form English results, last updated September 30, 2026, Whisper Large v3 Turbo averaged an 11.0% word error rate overall and 17.9% on its subset of recorded sociolinguistic interviews with speakers of African American English. Every model did worse on those interviews than on its overall average. Error rates of that size mean checking every quotation and every speaker change against the audio. Whisper models can also invent phrases: a 2024 study found whole hallucinated phrases or sentences in about 1% of the transcriptions it examined. Open ASR Leaderboard paper, long-form results, Careless Whisper study

If accuracy matters more to you than where the recording goes, consider ElevenLabs Scribe, our Rising Star. ElevenLabs, known for synthetic voices, launched its Scribe transcription model in February 2025. The version tested on the same leaderboard, Scribe v2, had the lowest error rates, 7.3% overall and 10.7% on the interview subset; its web app now uses the newer Scribe v3, which has not been independently tested. Its editor lets you move passages between speakers. Its lead among newcomers is narrow: Vocova, a cheaper hosted service with speaker editing and Word export, scores slightly lower on our criteria. It is a cloud service, and by default ElevenLabs may use what you upload to train its models: switch that off under Data use before your first upload, because the setting only applies from then on. Four one-hour interviews a month need the $22 Creator plan, or $220 billed yearly. On a Mac, also consider MacWhisper, which also works locally but lets you choose between Whisper and NVIDIA’s Parakeet models, has a fuller speaker editor and, by its developer’s account, runs much faster on Apple silicon; Pro costs €65 once. Scribe launch, ElevenLabs web app, ElevenLabs pricing, ElevenLabs privacy policy, MacWhisper

Disclosure: Practical Picks articles are researched and written with AI, currently Anthropic’s Claude. ElevenLabs offers a connection that lets Claude use its tools, MacWhisper can send a transcript to Claude for summaries if you add your own Anthropic key, and Vibe can summarize a transcript through the Claude API. We have no relationship with ElevenLabs, Good Snooze, the Vibe project or Anthropic over these products, and these AI connections carried no weight in our judgment.

Two free local apps complete the short list. Vibe exports Word, PDF, HTML, text and subtitle files but documents no way to correct speakers; Buzz adds speaker identification to a widely used transcriber but exports only text and subtitles. Rev, Descript, TurboScribe, Otter and several other hosted services scored lower on our criteria and are covered in “How we researched”. Our prices assume four one-hour interviews a month at normal rates checked November 12, 2026, before taxes and promotions. Four conditions would change our recommendation: accuracy matters more than privacy and cost (choose ElevenLabs Scribe), you use a Mac and want speed or fuller speaker editing (choose MacWhisper), you cannot wait hours or install software on your computer (choose a cloud service), or a person must verify the transcript (use a human transcription service such as Rev’s). Whatever you choose, try a short permissioned sample with names, numbers and a change of speaker first, keep the original recording and your corrected export, and check your institution’s rules before transcribing health, legal or other restricted interviews.

noScribe

Preferred Pick · interviews kept on your own computer

$0Free, open source · Windows or Apple silicon Mac

Choose noScribe when you have permission to transcribe an interview and want speaker-labelled text you can correct against the audio, without uploading the recording or paying anything.

Reasons to choose it

  • Runs entirely on your computer; works without a network connection, per its developer
  • Speaker detection, optional timestamps and a Word-ready file
  • Editor plays the audio from your cursor

What to watch

  • About 1–3 hours of processing per interview hour on an ordinary computer
  • Speakers corrected by editing labels or find and replace
  • Some Windows installs fail; no Intel Mac build

noScribe is a desktop app by Kai Dröge, a sociologist who teaches at Lucerne University of Applied Sciences and Arts, with help from other contributors. Version 0.8 came out on October 6, 2026 and was still the current version on November 12, under the GPL-3.0 open-source licence, for Windows 10 and 11, for macOS 14 or later on Apple silicon and, per its download page, for Linux as a downloadable executable. Georgetown University Library’s text-mining guide, updated September 22, 2026, lists it as a local transcription option and notes that it needs newer hardware and plenty of memory. The project’s code is public, its latest changes were committed on November 9, 2026 and its security page lists no published advisories. Project site, download and requirements, release notes, repository, Georgetown guide

It covers the whole job. Both bundled models are versions of Whisper V3 Turbo, and pyannote separates the speakers; its usage guide advises entering the number of speakers, because detection sometimes finds more people than were present. You can ask for a timestamp every 60 seconds and at each change of speaker, mark overlapping speech (experimental) and choose whether to keep filler words. The default output is HTML that opens in Word or LibreOffice and imports into qualitative-analysis tools such as MAXQDA; plain text and WebVTT subtitles are alternatives. Usage guide, FAQ, models

Read the full noScribe analysis and sources

Correction happens in the separate noScribe Editor. A keyboard shortcut plays the audio from the cursor and the selection follows the playback; version 0.8 added a seek bar and custom speaker names. You rename a speaker throughout with find and replace and move a passage to another speaker by editing its label, which is workable for a two-person interview but slower than the reassign controls in MacWhisper, Rev or ElevenLabs. The developer’s own guide warns about extra speakers, misspelled names and Whisper’s occasional invented phrases, and says languages such as Persian, Punjabi and Tamil are transcribed much less accurately than Western languages. Usage guide, release notes

Choose something else if you are on deadline or on older hardware. The developer estimates one to three hours of processing per hour of interview, longer on older machines; a supported NVIDIA graphics card with at least 6 GB of memory shortens that considerably. Intel Macs stopped at version 0.6, and Windows warns about an unknown publisher during installation. Open bug reports describe installations or jobs that fail: one after upgrading to 0.8 on Windows 11 that cannot load required libraries (filed October 14, 2026), one at the speaker-detection step on Windows 10 (May 28, 2026) and a newer one in which jobs keep failing with an “Errno 2” error in 0.8. The developer could not reproduce the first and suggested updating Microsoft’s Visual C++ libraries. Try a short file before you depend on it, and download only from noscribe.de or GitHub: the developer warns that a lookalike .ai domain sells paid transcription services unconnected to the project. Download and requirements, issue 373, issue 303, issues list, project site

Nothing leaves your computer and there is no account, so your recordings and transcripts are only as safe as your own disk and backups. The project depends on one main developer, who has written most of the code, and on donations. The open-source licence means others could continue it, and transcripts are ordinary HTML or text files you keep whatever happens to the project. FAQ, repository

Used noScribe? Share your experience to help other readers choose.

ElevenLabs Scribe

Rising Star · a newcomer worth considering

$22/moCreator · about 6 hours of transcription credits · $220 billed yearly

Why Rising Star? ElevenLabs entered transcription in February 2025. Its Scribe v2 model had the lowest long-form error rate in an independent public benchmark; its web app now uses the newer Scribe v3, which has not been independently tested. Its cloud processing and default training terms are the limits. Its lead over Vocova, a cheaper hosted newcomer, is narrow. This is not our broader Ready to Rock endorsement.

Consider ElevenLabs Scribe when transcription accuracy matters more than where the recording is processed, and switch off model training before your first upload.

Reasons to choose it

  • Scribe v2 had the lowest long-form error rate on the Open ASR Leaderboard
  • Editor moves passages between speakers and adjusts timing
  • More than 90 languages; up to 32 speakers

What to watch

  • Trains on uploads by default; opt-out applies only afterwards
  • Retention of web-app uploads not documented
  • Proposed class action over voice-model training (allegations)

ElevenLabs entered transcription in February 2025 with Scribe, its first speech-to-text model, available in its web dashboard and its programming interface from the start. Its web app accepts audio or video files up to 3 GB, and its models page lists up to 32 speakers per file. We name it Rising Star because it is a new entrant to transcription with a substantiated advantage on the part of the job that matters most, not because of the company’s size. This is not our broader Ready to Rock endorsement: we found little independent evidence of its use for interviews beyond benchmarks. Scribe launch, web app guide, models

Among the 27 systems in the Open ASR Leaderboard’s long-form English results, updated September 30, 2026, Scribe v2 had the lowest average word error rate, 7.3%, against 9.5% for Rev AI Fusion, 10.7% for NVIDIA Parakeet TDT 0.6B v3 and 11.2% for Whisper Large v3. On the recorded-interview subset the same four scored 10.7%, 12.6%, 15.6% and 18.9%. Artificial Analysis’s separate index placed Scribe v2 third of its listed models, at 2.2%, when we checked on October 19, 2026. ElevenLabs’ web app guide says Scribe v3 is the default model on its website. It describes v3 as its most accurate model for recorded audio and says it makes 27% fewer word errors than v2 on its own benchmark; we found no independent test of v3, so the independent results above describe the previous model, not necessarily the one that transcribes your upload. Both independent sources measure models through programming interfaces, not the web app, and the leaderboard paper’s authors include people from NVIDIA, Mistral AI and OpenAI, whose models it ranks, though its code and data are public. Better scores reduce correction work; they do not remove it. Long-form results, Open ASR Leaderboard paper, Artificial Analysis index, ElevenLabs models

Read the full ElevenLabs Scribe analysis and sources

The transcript editor is documented in detail. You edit text directly, rename speakers, move one segment or many to another speaker, adjust a segment’s start and end times and re-align words after editing. Exports listed in the documentation are plain text, HTML, JSON, SRT and VTT; Word is not listed, so open the HTML export in Word if you need a document. A human transcript review service, from $2 a minute, is in beta for selected users. Transcript editor

You pay in credits shared with ElevenLabs’ other tools, and transcription uses 330 credits a minute. By our arithmetic, the free plan’s 10,000 credits cover about 30 minutes a month without a commercial licence, Starter at $6 a month (30,000 credits) covers about an hour and a half, and Creator at $22 a month (121,000 credits, $11 for the first month) covers about six hours, enough for four one-hour interviews. Annual billing costs ten months of the monthly price, $220 a year for Creator. ElevenLabs’ terms say subscriptions renew at its then-current fee and that all sales are final, though it may offer refunds at its discretion; you can cancel at any time and the plan then runs to the end of the billing period. Pricing, terms

Data handling is the main reason to hesitate. ElevenLabs’ privacy policy, updated October 9, 2026, says it uses inputs and outputs, which can include voice recordings, to train its models; you can switch this off under Data use, but only for what you upload afterwards. Transcripts can be deleted through its programming interface, although ElevenLabs says debugging and moderation logs may keep related data, and its zero-retention mode is for selected Enterprise customers and never covers the web app. On May 11, 2026, journalists and voice actors filed a proposed class action in federal court in Illinois alleging that ElevenLabs trained voice models on their publicly distributed recordings without the notice and consent Illinois biometric law requires. Those are allegations, not findings, and they concern public recordings rather than transcription uploads. Do not use it for interviews you have promised to keep off third-party servers. Privacy policy, data-use setting, deleting transcripts, zero-retention mode, complaint, Amer v. Eleven Labs

Used ElevenLabs Scribe? Share your experience to help other readers choose.

MacWhisper

Also consider · Mac users

€65 oncePro licence · free edition available

Choose MacWhisper on an Apple silicon Mac when you want local transcription with faster processing, a choice of models and fuller speaker editing than noScribe.

Reasons to choose it

  • Local by default; choose Whisper or Parakeet models
  • Pro renames, merges and reassigns speakers
  • Word, PDF, text and subtitle exports with timecodes

What to watch

  • Mac only; speaker recognition needs Pro
  • Cloud options send audio or text to other companies
  • Privacy policy from 2024 predates its cloud features

MacWhisper is made by Jordi Bruin’s Good Snooze in the Netherlands. The direct version, sold through Gumroad, has a free edition and a Pro licence with lifetime updates: €65 on the Gumroad checkout page and €64 on the developer’s own site, so we use the higher figure. The developer offers 25% off to journalists, students and nonprofits on request and a refund within seven days if you are unhappy. It needs macOS 14 or later and recommends an Apple silicon Mac. The App Store version, called Whisper Transcription, is sold by subscription with fewer features, so check which one you are buying. Sales page, site, editions

Pro adds the parts an interviewer needs: automatic speaker recognition with supported local models on Apple silicon Macs, or through the ElevenLabs and Deepgram cloud services, plus batch transcription and NVIDIA’s Parakeet models. The free edition lets you add up to two speakers by hand. In the editor you can rename and merge speakers and reassign a segment with a number key. Exports include Word, PDF, plain text, Markdown, HTML, CSV and SRT and VTT subtitles, with timecodes. The developer claims transcription at up to about 30 times real time with Whisper, and faster with Parakeet; we did not measure it. For English interviews, Parakeet v3 had a lower error rate than Whisper Large v3 on the leaderboard’s interview subset, 15.6% against 18.9%. Sales page, speaker recognition, long-form results

Read the full MacWhisper analysis and sources

Transcription runs on your Mac by default. Choosing a cloud transcription provider sends the audio to that provider, and translation and AI prompts send the transcript to the translation or AI service you set up, unless you use a local model. The linked privacy policy, dated February 14, 2024, still says no data leaves your device; we rely on the newer feature documentation instead, and the mismatch is a reason to check your settings before you transcribe sensitive recordings. Local and cloud processing, AI providers, privacy policy

Choose noScribe instead if you use Windows or want to pay nothing. The personal licence has a soft limit of three devices, and since December 2025 the developer has disabled licences automatically when it detects resale or overuse, so remove a licence from an old Mac before moving to a new one. Licensing

Used MacWhisper? Share your experience to help other readers choose.

Vibe

Free local app · many export formats

$0Free, open source · Windows, Mac or Linux

Consider Vibe when you want a free local app that exports Word, PDF, HTML and subtitle files, and you can fix speaker labels in another tool.

Reasons to choose it

  • Runs on your computer, per its README
  • Exports DOCX, PDF, HTML, TXT, SRT and VTT
  • Speaker diarization listed; Windows, macOS and Linux

What to watch

  • No way to correct speaker labels is documented
  • No transcript editor described in the pages we read
  • Pull requests limited to a few collaborators

Vibe is an open-source (MIT) desktop transcription app published under the GitHub account thewh1teagle, with 49 contributors listed. It runs on macOS, Windows and Linux; its latest release, version 3.3.0, came out on October 29, 2026, and its repository lists 102 releases. The project is not new: files in its repository are dated January 2024. We read its repository page and README, not its website, privacy policy or installers, so what follows is the project’s own description. Repository and README

The README lists transcription of audio and video files on your own device, batch transcription, Whisper, Parakeet TDT v3 and Nemotron models, translation to English, a stable-timestamps mode for subtitle timing, a command-line interface and an HTTP interface. “Speaker diarization” is listed, and recent changes in the repository replace its diarization model with NVIDIA’s Nemotron-3-Diarization, which handles up to 8 speakers, and split transcript segments at speaker changes. Exports are SRT, VTT, TXT, HTML, PDF, JSON and DOCX, so it is the one local app here that lists a Word format directly. For English interviews, the leaderboard gives Parakeet v3 15.6% on its interview subset; Vibe does not say which model it uses by default. Repository and README, long-form results

Read the full Vibe analysis and sources

Choose noScribe instead if you need to correct and check the transcript against the audio. We found no description of a transcript editor, of how to reassign or rename speakers, or of timestamp and speaker options in the document exports. The README says transcription happens on your device and no data leaves it; optional features, such as summaries through the Claude API or Ollama and model downloads, use other services. The repository says pull requests are limited to a few collaborators. Repository and README

Used Vibe? Share your experience to help other readers choose.

Buzz

Free local app · text and subtitle files

$0Free, open source · Windows, Mac or Linux

Consider Buzz when you want a free, widely used local transcriber with speaker identification and you only need text or subtitle files.

Reasons to choose it

  • Runs on your computer, per its README
  • Speaker identification and a known-speaker-count setting
  • Viewer with search, playback controls and speed adjustment

What to watch

  • Exports only TXT, SRT and VTT
  • Speaker identification is new and its correction tools are not documented
  • Windows installer is unsigned; Intel Mac support ends after version 1.4.5

Buzz is an open-source (MIT) desktop app that transcribes and translates audio on your own computer using OpenAI’s Whisper, published under the GitHub account chidiwilliams, with 53 contributors listed. It runs on Windows, Linux and, on Apple silicon, macOS; Intel Macs stopped at version 1.4.5. Its latest release is 1.4.5 from August 23, 2026, and the latest commit we saw was on November 9, 2026. Its Windows installer is unsigned, so Windows warns about it. The project is not new: files in its repository are dated from September 2022. We read its repository page and README, not its documentation site or installers. Repository and README

The README lists speaker identification in transcribed media, speech separation before transcription for noisy audio, a transcription viewer with search, playback controls and speed adjustment, a watch folder, a command-line interface and plugins. Recent commits add a known-speaker-count setting and align long transcripts in chunks when identifying speakers, so the speaker feature is still changing. Exports are TXT, SRT and VTT only; there is no Word-openable format, so you would paste text into a document yourself. It does not say which Whisper version it uses by default. Repository and README

Read the full Buzz analysis and sources

Choose it when you want a free, local tool and only need text or subtitle files. We found no description of how to correct speaker labels in the viewer. The README says it transcribes on your computer, and we read nothing about where optional plugins send data. Repository and README

Used Buzz? Share your experience to help other readers choose.

Which product are we missing?

Know a product that belongs in this article? Tell us what it does differently and why it deserves a closer look. Newcomers welcome.

Readers and product teams can suggest candidates. Suggestions inform our research; they do not guarantee inclusion or a recommendation.

Suggest a product

Opens an email draft with this article attached as a link. Nothing is sent until you send it.

No email app? Send the product name, website, and this article’s title to admin@practicalpicks.guide.

What people are saying

User reports point to the same conditions as the documentation. On noScribe’s GitHub tracker, open reports from 2026 describe Windows installations that fail before transcription starts and version 0.8 jobs that keep failing. An October 2025 MacWhisper discussion complains about speaker assignment and exports in version 13.0.2 (read September 6, 2026), and MacWhisper’s own sales page showed 4.5 stars from 2,784 ratings on October 19, 2026. We did not read user discussions of Vibe or Buzz. These are self-selected reports, not representative samples, and none of them measures accuracy. noScribe issue 373, issue 303, issues list, MacWhisper discussion, MacWhisper sales page

One court case matters for readers who care where voices end up: Amer v. Eleven Labs Inc. (Northern District of Illinois) was filed on May 11, 2026 and contains allegations that have not been proven. ElevenLabs complaint

How we researched

This is public-source research. We did not upload recordings, create accounts, install software, buy anything or measure accuracy ourselves. No vendor supplied access, payment or input, and we use no affiliate links. We first compared Otter, Descript and MacWhisper on September 6, 2026. On October 19, 2026 we rechecked every material claim, added four tools and read vendor documentation, help-center articles, privacy policies and terms, peer-reviewed studies, public benchmarks, court records, project issue trackers and review profiles. On November 12, 2026 we rechecked prices and plan limits for every tool, rechecked noScribe’s releases, security page and open issues, reread ElevenLabs’ pricing, terms and privacy policy, read the repositories of two free local apps, Vibe and Buzz, that we had not evaluated before, and read the product, pricing and policy pages of three recent launches.

We wrote this for someone with permission to transcribe an ordinary recorded interview, about four one-hour recordings a month, who needs the right speaker on each passage, timestamps and a document that opens elsewhere. We fixed five criteria and their weights before scoring: public accuracy evidence (25%), speaker detection and correction (20%), document export and timestamps (10%), handling of the recording and transcript (25%) and cost for that workload (20%), counted over three years. The Preferred Pick also had to run on Windows and Mac. noScribe scored highest, narrowly ahead of MacWhisper, which runs only on Macs, and ahead of Vibe and Buzz; ElevenLabs Scribe scored lower on those weights. Giving accuracy 40% of the weight still leaves noScribe and MacWhisper ahead; ElevenLabs Scribe comes out ahead of every other tool only when accuracy carries about half the weight, which is why we name it for readers who put accuracy first. Adding setup and processing time as a criterion favors MacWhisper on a Mac.

Rev, Descript, TurboScribe and Otter, which we evaluated in full in October, scored below the top four and no longer have their own sections. Rev (Essentials $25.49 a month billed yearly) has the most complete hosted transcript editor and sells human transcription from $1.99 a minute, but its terms let it train its speech models on your content; Descript (Hobbyist $16 a month billed yearly) suits an interview that will become an edited podcast or video, which none of the five tools above does; TurboScribe (Unlimited $10 a month, $120 billed yearly) is the cheapest unlimited cloud service, but its own pages blocked our automated check in October, so we relied on archived and search-indexed copies then, and its pricing page was readable on November 12; Otter (Pro $8.33 a month billed yearly) is built for meetings, trains on uploads under its privacy policy and allows ten imports a month on Pro. These descriptions rest on the October reading of their documentation. We priced five hosted services we had not evaluated in October from their pricing pages only: Happy Scribe (Pro $19 a month billed yearly, 600 uploaded minutes a month), Sonix (Core $275 a year for 60 hours), Notta (Pro $97.99 a year), Transkriptor (Pro $99.99 a year) and Amberscript (Starter €190 a year for 60 hours). We also looked at Ekhos, a Windows-only local transcriber at $9 a month. We read no privacy policies or accuracy evidence for these six, and none came close to the top four on the criteria we could check. We could not read a Trint price (its pricing page returned an error) or Microsoft’s Word transcription feature page, and we did not read the upload features of meeting assistants such as Fireflies, so we have not scored them; developer interfaces from AssemblyAI, Deepgram and Speechmatics are not upload tools for this reader.

We also searched for new entrants on November 12, 2026 through GitHub topic pages, Hacker News launch posts and developer repositories. ElevenLabs Scribe still meets our Rising Star threshold: its first public release was in February 2025, it runs in a browser on any computer, its editor and exports are documented, and its operator is identified. Vocova, a hosted service launched in 2026 and run by a named UK company, also meets the conditions: it labels speakers, lets you edit speakers and timestamps and exports Word and subtitle files, and its Plus plan, which covers four interviews a month, costs $90 billed yearly. It scored slightly below ElevenLabs Scribe on our criteria, so ElevenLabs Scribe keeps the award, but the lead is narrow. LymeScribe, a local Mac and Windows app with a one-time $39 to $59 unlock, labels speakers, but on Windows only with an NVIDIA graphics card, and we could not confirm its exports, timestamps or transcript correction in the pages we read, so it fails our mandatory test for a recommendation and cannot take the award. Harku, a hosted interview tool, names no operator on the pages we could read, which fails our safety floor for a service that stores your recordings. We also screened a Windows-only local transcriber launched in 2025, several other 2026 launches (a hosted service that publishes no speaker labelling, price or operator, a voice-dictation tool with a command-line transcription option, a recorder-bound desktop app and a transcript-based video editor) and a number of single-platform and meeting-capture tools, and none met the conditions of working on both Windows and macOS, doing the core task in public sources and having an identified operator. Vibe and Buzz, the two free local apps now in our top four, are more than two years old, so neither can be a Rising Star. We will look again at the next review.

ToolPublic sourcesWhat remains uncertain
noScribeProject site, usage guide, FAQ, model notes, download page, release notes, issue tracker, security page, Georgetown library guide.Accuracy and speaker separation of its exact pipeline, how often installs fail, processing time on your computer.
ElevenLabs ScribeLaunch post, web app and editor guides, models page, pricing, terms, October 2026 privacy policy, data-use help, deletion reference, Illinois complaint.Retention of web-app uploads, Word export, which model version your upload uses, outcome of the lawsuit.
MacWhisperSales page, site, local and cloud processing, speaker, edition, AI-provider and licensing pages, 2024 privacy policy and terms.Which features the free edition keeps, behavior when cloud options are configured, speed on your Mac.
VibeRepository page and README.How to correct speakers, default model, document export options, privacy policy, behavior of optional network features.
BuzzRepository page and README.How to correct speakers, default model, behavior of plugins, speaker feature maturity.

None of this establishes accuracy on your recordings, correct speaker labels or deletion in practice. Next substantive review is due December 18, 2026, or sooner if a price, training term or court ruling changes. Found a changed term? Contact the editors.

Sources and evidence

Sources were checked October 19, 2026 unless another date is given; prices and plan limits were checked November 12, 2026. Vendor material establishes documented offers and terms, not transcript quality. Benchmarks measure speech models on public test sets, not these products on your recordings.

This is a researched comparison, not a product review. Method: evidence standard, editorial picks.