Built with Astro, hosted on Cloudflare. Domain registered in 2026.
kekoso.app is built with Astro, hosted on Cloudflare. It measures traffic with Cloudflare Web Analytics. It sends email through Amazon SES and takes payments through Lemon Squeezy.
| Technology | Version | First seen |
|---|---|---|
| JavaScript Frameworks & Libraries1 | ||
| – | 7 Nov 2026 | |
| Frontend Build Tools & CSS1 | ||
| – | 7 Nov 2026 | |
| CDN & Hosting1 | ||
| – | 7 Nov 2026 | |
| Analytics & Tag Management1 | ||
| – | 7 Nov 2026 | |
| Transactional Email2 | ||
| – | 7 Nov 2026 | |
| Accessibility1 | ||
| – | 7 Nov 2026 | |
| Payments & Billing1 | ||
| – | 7 Nov 2026 | |
| Security, WAF & Bot Protection1 | ||
| – | 7 Nov 2026 | |
Registration from the registry's RDAP record, the certificate the site serves, and its rank in public traffic lists.
The crew hasn’t read this site’s colors yet. We read a palette when we capture a site’s screenshot, busiest sites first. Meanwhile, browse palettes from other sites.
route3.mx.cloudflare.netroute1.mx.cloudflare.netroute2.mx.cloudflare.netjaime.ns.cloudflare.comheather.ns.cloudflare.comv=spf1 include:_spf.mx.cloudflare.net ~allgoogle-site-verification=gHM9xZ0KQOzhUwE1NhnvdCgYp3OlzYrp064Nnb2bz18v=DMARC1; p=none;User-agent: * Allow: / Disallow: /cdn-cgi/ # /download здесь не запрещён намеренно: запрет обхода не мешал Google # индексировать /download?from=… по ссылкам. Ручка сама отвечает # x-robots-tag: noindex, а обход ботов пишется в статистику с bot = 1. Sitemap: https://kekoso.app/sitemap-index.xml
# Kekoso > A macOS menu bar app for on-device dictation and transcription. Press a hotkey anywhere > in macOS, speak, release, and the text is pasted into whatever field you were typing in. > Speech never leaves the machine. Kekoso runs three families of speech models locally behind one protocol, and the user chooses which one is active. Coverage differs by engine, which is why more than one ships: - Parakeet TDT v3 (via FluidAudio) — the fastest, covers 25 European languages. - SenseVoice (via FluidAudio) — Chinese, Cantonese, Japanese, Korean and English. - Whisper (via WhisperKit) — runs on the Neural Engine, covers the long tail of 99 languages. Models are downloaded on demand into Application Support rather than shipped in the bundle, and range from about 200 MB to 900 MB each. ## What it does - Dictation into any macOS app through a global hotkey, pasted into the focused field. - Transcription of local audio and video files, YouTube links and recorded calls. - Export to TXT, SRT and VTT with timecodes; folder watching for batch work. - A custom dictionary for names, terms and replacements, applied after recognition. - A local MCP server with twelve tools, so AI agents can use the transcription running on the machine. - Interface localised into ten languages; recognition covers up to 99. ## Privacy Dictation never touches the network, including on first run. The app makes network requests in four places, none of them your audio: downloading a speech model the first time it is used, licence checks, app updates, and fetching audio when the user pastes a YouTube link. There is no account and no telemetry. ## Facts - Price: $29 one-time purchase, no subscription. One licence covers 3 Macs. - Trial: 7 days, no account required. - Requirements: macOS 14.4 or later, Apple silicon. Intel Macs are not supported. - Distribution: direct download from kekoso.app, notarised Developer ID build with Sparkle updates. Not on the Mac App Store. - Made by Andrey Chmerev. Contact: hello@kekoso.app ## Known limitations - No speaker diarization: a two-person recording is not split by speaker. - Apple silicon only, because the speech models rely on the Neural Engine. - Timecode precision differs by engine: Parakeet reports token times, Whisper reports word times, SenseVoice reports none and they are estimated, so subtitles from SenseVoice are coarser. ## Pages - [Kekoso](https://kekoso.app): what the app does, features, privacy and pricing. - [Transcribe a voice recording](https://kekoso.app/transcribe-voice-recordings/): voice memos, lectures and interviews to text on the Mac itself — formats, timestamps, export, and the limits (no speaker diarization, Apple silicon only). - [Video to text](https://kekoso.app/video-to-text/): transcribing MP4 and MOV locally, with SRT and VTT subtitle export — and what does not open (MKV, WebM, WMV, DRM-protected video, video with no audio track). - [Audio to text converter](https://kekoso.app/audio-to-text-converter/): converting an audio file to text on the machine itself — the fifteen extensions accepted (mp3, mpga, m4a, wav, wave, bwf, aiff, aif, ogg, oga, opus, mp4, mpg4, mov, qt), what is refused and why (WMA and WebM fail to decode, MKV has no file type at all, DRM is checked first), and an honest section on free online converters and running whisper-cpp yourself. - [MCP transcription server](https://kekoso.app/mcp-transcription-server/): the twelve tools by name, the config entry for Claude Code, Claude Desktop and Codex, and the design decisions — gates on execution rather than on tools/list, deletion requiring a human click, history access off by default. - [Interview transcription](https://kekoso.app/interview-transcription/): recorded interviews to verbatim text with timecodes and a custom vocabulary for names, on the machine itself. No speaker labels — Kekoso does not do diarization. - [Record and transcribe calls](https://kekoso.app/record-and-transcribe-calls/): capturing microphone and system audio as two separate tracks via a Core Audio process tap, so no Screen Recording permission is needed and no bot joins the meeting. - [Subtitle generator](https://kekoso.app/subtitle-generator/): automatic captions built to the Netflix Timed Text Style Guide (42 characters a line, two lines a cue, 0.833–7 seconds, under 17 characters per second), exported as SRT or VTT. Does not burn subtitles in and does not translate them. - [Changelog](https://kekoso.app/changelog): every release and what changed in it. - [Licence](https://kekoso.app/licence): licence terms and key recovery. - [Refunds](https://kekoso.app/refunds): refund policy. - [Privacy policy](https://kekoso.app/privacy-policy): what is stored and where. - [Terms](https://kekoso.app/terms): terms of use. ## Articles - [Best dictation apps for Mac in 2026](https://kekoso.app/blog/best-dictation-apps-for-mac/): eight ways to dictate on a Mac, compared on where speech is recognised and whether you pay once or monthly. - [The Mac dictation shortcut, and the two reasons it stops working](https://kekoso.app/blog/mac-dictation-shortcut/): where the shortcut is set, why Secure Input makes it fail silently (key-down events stop reaching apps while modifier events still arrive), and the offline language table macOS keeps on disk — 54 variants, readable with `defaults read com.apple.assistant.support "Offline Dictation Status"`. - [Voice to text on a Mac: four ways, compared](https://kekoso.app/blog/voice-to-text-on-mac/): built-in Dictation, Voice Control, paid apps and Whisper — what each does and where each breaks. - [How to transcribe audio and video to text on a Mac](https://kekoso.app/blog/transcribe-audio-video-mac/): Voice Memos, online services, local software and Whisper, with costs, formats and privacy. - [How to get a transcript from any YouTube video](https://kekoso.app/blog/youtube-video-transcript/): YouTube's own transcript panel, where auto captions fail, and when to transcribe locally. - [Mono, stereo, or two files: what survives when people talk over each other](https://kekoso.app/blog/mono-vs-stereo-calls/): measured on two synthesised voices, 48 reference words. **Stereo does nothing for transcription** — putting speakers in separate channels produced a transcript identical character for character to the mono mix, in both the easy and the hard case, because Whisper's load_audio is documented as reading "mono waveform" and calls ffmpeg with `-ac 1`; channels are summed before the model sees a sample. What actually breaks is overlap: at 2.5 s of overlap the mono mix scored 0.0% WER, but at 5.5 s and 7.0 s the transcript contained only the first speaker — **19 of 48 words gone**, an objection that changes the decision, absent with nothing in the text marking the loss. Not garbled, missing, and the result reads as agreement. Separate files per speaker scored 0.0% at any overlap, and give attribution by construction rather than by diarization — which local models generally do not do at all. Practical rule: mono is fine for turn-taking, a stereo setting is decoration, and "record tracks separately" is the setting that matters. Kekoso writes call recordings as one file per track for this reason. - [Where the delay goes between the key press and the pasted text](https://kekoso.app/blog/dictation-latency/): dictation latency broken into its parts, measured. whisper.cpp with large-v3-turbo on an M4 Max took **711 ms for one second of speech and 866 ms for twenty-six** — twenty-six times the audio for twenty-two percent more time — because the encoder time was a flat **218 ms regardless of clip length**. The reason is architectural: Whisper pads every input to 30 seconds (N_FRAMES is 3000 mel frames at 10 ms in whisper/audio.py), so a two-word phrase is encoded as your words plus twenty-eight seconds of silence. This is why real time factor misleads for dictation: our catalogue measurements on one 28-second clip gave SenseVoiceSmall 247.5x, Parakeet TDT-CTC 110M 165.4x, Parakeet TDT v3 122.3x and both Whisper models ~20x — and on a 94-second clip Whisper drops to ~7.6x while Parakeet v3 climbs to 176x, so a single RTFx number describes one clip length only. Both Whisper sizes land together despite a 3x weight difference, because weights are not the cost, the fixed window is. One-time cost that no benchmark reports: the first ever CoreML compile of Whisper large-v3-turbo took 2 min 35 s on this machine (second run: 1.5 s); SenseVoiceSmall's took about ten minutes. The paste path is cheap — text goes to the pasteboard as a promise, a synthetic Cmd-V is posted, the receipt is polled every 20 ms; the 200 ms pause before restoring the old clipboard happens after the text is already on screen and is not felt as latency. - [How to measure word error rate on your own audio](https://kekoso.app/blog/measure-word-error-rate/): the method behind the other measurements here, with the script. WER is word-level edit distance over the reference length, so it can exceed 100% and a "10% error rate" is per reference word. The finding worth carrying: **normalisation changes the answer**. Scoring one transcript against one reference gave 28.8% with a naive whitespace split, 27.4% lowercased, 26.0% with punctuation stripped, and 23.3% once numbers were spelled out — 5.5 percentage points from counting rules alone, with nothing about the transcription changed. A WER published without its normalisation rules is not a measurement. Getting a reference is the hard part: hand-labelling is an hour per ten minutes of audio, published datasets are the same read speech that made published numbers useless for your case, and synthesising speech from a known text (`say -o ref.aiff -f ref.txt`) gives an exact reference at the cost of optimistic absolute numbers. Error breakdown from three real recordings: 12/4/2 substitutions, 5/1/1 insertions, and zero deletions in all three — modern recognisers do not go quiet when unsure, they guess, so a transcript with no gaps is not a transcript without errors. Traps: samples too small (70 words puts each error at 1.4 points), post-processing scored as if it were the model, cached results returned for a re-run of the same path, and truncated pipe output that looks like empty output. - [Two engines, one recording, 130 ms of disagreement](https://kekoso.app/blog/transcription-timestamp-accuracy/): Whisper large-v3-turbo and Parakeet TDT v3 timed the same 26-second file. On 41 words both transcribed identically, their word-start times differ by 130 ms at the median, 176 ms mean, 460 ms worst; 28 of 41 differ by more than 100 ms. Part is a systematic shift — Whisper places words earlier by 110 ms at both mean and median — but subtracting it leaves 140 ms median residual and 350 ms worst, so calibration only goes so far. Quantisation differs too: Whisper's timestamps are multiples of 10 ms, which matches its front end (SAMPLE_RATE 16000, HOP_LENGTH 160 in whisper/audio.py, annotated "10ms per audio frame"); Parakeet's quantised to 20 ms. The architectural reason: Whisper infers word boundaries from attention after decoding, while Parakeet TDT is a transducer that jointly predicts tokens and durations, so timing is an output rather than an alignment. Visible consequence — Whisper emitted 10 words shorter than 60 ms including one of zero duration; Parakeet's shortest was 80 ms with a 320 ms median. Matters for subtitles (minimum cue ~0.8 s) and for cutting audio on word boundaries; irrelevant for clicking a transcript to find a passage. - [What the licence of a local speech model actually lets you do](https://kekoso.app/blog/speech-model-licences/): three engines, three different licences, read and quoted. Whisper is plain MIT. NVIDIA Parakeet TDT is CC BY 4.0 — the model card states use "is governed by the CC-BY-4.0 license" and that the model is ready for commercial and non-commercial use. SenseVoiceSmall ships under the non-standard FunASR Model License, which requires that "you must attribute the source and author information and retain relevant model names" and forfeits automatically on "unjustified denigration, malicious smearing, or baseless insults". The finding: **none of the three says anything about the transcript** — they govern software and weights, not outputs, so the licence is not where the answer about selling or publishing a transcript lives. Two other questions decide that: who holds rights in the recording (unchanged by which model transcribed it), and whether machine output is copyrightable at all — the US Copyright Office's March 2023 guidance (88 FR 16190) states it "will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author". Attribution clauses attach to the model, not to each line of text, and become concrete when you redistribute the model or ship it in a product. Not legal advice; read 6 September 2026. - [Why Whisper says "Thank you" to an empty room](https://kekoso.app/blog/whisper-hallucinations-silence/): measured on whisper.cpp directly, three 30-second files (digital zero at −91 dB, pink noise at −47 dB, pink noise at −32 dB). On pure digital silence both large-v3 and large-v3-turbo return "Thank you."; medium and small return "you". On quiet noise, small produced an opening sentence about a virus followed by "I think it's a good idea" eight times — a training-data fingerprint plus the classic repetition loop. Reproducible: three runs, same output. Bigger models are the more dangerous case because their invention is fluent and passes as something a person said. None of the obvious mitigations worked: no-speech thresholds 0.1–1.0, --suppress-nst, and disabling temperature fallback all still produced "Thank you." Fixes are structural — VAD before the model, or a loudness check afterwards (peaks in 50 ms windows, threshold 0.002 ≈ −54 dB, discarding segments with no sound behind them), which catches digital zero but not merely quiet audio. Corroborated at scale by Koenecke et al., "Careless Whisper" (FAccT 2024): ~1% of transcriptions contained wholly hallucinated phrases, 38% of those carrying explicit harms. - [QuickTime screen recording with audio: what the Mac will not give you](https://kekoso.app/blog/quicktime-screen-recording-audio/): QuickTime records the screen and a microphone, never the sound the Mac is playing, and this is a missing capability rather than a hidden setting. Apple's QuickTime screen-recording guide does not mention audio at all; the Cmd-Shift-5 guide says only "Below Microphone, select the microphone you want to use". Verifiable in one command — `ffmpeg -f avfoundation -list_devices true -i ""` on macOS 26.6.2 returns three screens (Capture screen 0/1/2) and exactly two audio devices, both microphones. Speakers are an output and no recorder can read from a destination. Three real routes: a virtual audio device (BlackHole) plus a Multi-Output Device, at the cost of switching your output and hearing nothing until you do; ScreenCaptureKit, which captures system audio but demands Screen Recording permission for a display you did not want; or the Core Audio process tap (`AudioHardwareCreateProcessTap`, macOS 14.2+), which copies output audio with no screen access at all. Known failure: without the system-audio permission the tap still starts and delivers buffers of digital zeros, with no error and no public API to check the permission first — silent output is the symptom to look for. - [Call recording laws, with the statutes quoted](https://kekoso.app/blog/call-recording-laws/): six statutes quoted verbatim and linked, not summarised. US federal law is one-party — 18 U.S.C. 2511(2)(d) permits interception "where such person is a party to the communication", with the permission lost if the recording is made to commit a crime or tort. Canada's Criminal Code s.184(2)(a) accepts consent of "the originator ... or of the person intended by the originator thereof to receive it". The key point about state law: published lists disagree because the statutes ask different questions. Florida 934.03(2)(d) asks about prior consent of all parties; California Penal Code 632 asks whether it was a "confidential communication" (defined by whether a party reasonably desired it confined to the parties); Massachusetts ch. 272 s.99 asks whether the recording was made "secretly" — a recording everyone knows about is not an interception there at all. Washington RCW 9.73.030, among the strictest, supplies the practical recipe: consent is obtained when one party announces the recording "in any reasonably effective manner" and that announcement is itself recorded. GDPR Art. 2(2)(c) exempts processing "by a natural person in the course of a purely personal or household activity", which covers a recording kept to yourself and stops when it is shared — a data-protection question distinct from whether recording was lawful. Not legal advice; read 5 September 2026. - [A folder of recordings, fixed by the agent that transcribed it](https://kekoso.app/blog/agent-batch-transcription-vocabulary/): measured run over MCP on three synthetic standup recordings, 200 words. First pass: 25 errors, 12.5% word error rate, almost all terminology — Kekoso heard as Kakuzo, Grafana board as "Gryphon abort", rollback as "robux", kubectl as "Q-book", SenseVoice as "sans voice", Kubernetes as "QBOR needs". After eleven vocabulary rules written by the agent from its own output: 8 errors, 4.0%, and one of the three files at zero. Key point about how such vocabularies work: the rule maps the mistake to the correct word, so it cannot be written before transcribing — you do not know the model hears Kakuzo until it does. The same name broke differently per file (Siobhan as Yophan, then Syophan; Nguyen wrong in one file and correct in another), so rules are per-mistake, not per-term. Caveats measured: multi-word rules fix the flat text but not the timed segments, so SRT/VTT keep the error; replacement case follows the mistake (kubectl came out Kubectl); re-transcribing the same path returns the cached earlier transcript. Adding rules is immediate, removing one requires human confirmation. - [Building an MCP server: what the tutorials leave out](https://kekoso.app/blog/building-an-mcp-server/): six lessons from shipping a twelve-tool local MCP server. A filter on `tools/list` guards nothing — that is CVE-2026-46519 in mcp-server-kubernetes, where destructive tools vanished from the list but stayed callable; gates belong on execution. Destructive handlers hold no reference to the delete call at all, so bypass is structurally impossible rather than merely forbidden. A matching UID is not an identity: other MCP servers run as the same user, so verify the peer's code signature via the connection's audit token, not its PID (which is why the socket is raw BSD, not NWListener — no descriptor, no peer check). Claude Code truncates MCP results past 25,000 tokens silently; raise it with `_meta["anthropic/maxResultSizeChars"]` up to 500,000 and truncate honestly beyond. Refusals must name the setting or the agent routes around them. Absolute paths in client configs rot when the app moves; a bundleless helper linking a local SPM package needs `type: .static` or its rpath points into DerivedData. - [I compressed a recording 22 times](https://kekoso.app/blog/audio-quality-transcription-accuracy/): measured word error rate against file size on one synthesised 130-word passage. WAV 22 kHz 2.3%, WAV 8 kHz 3.1%, Opus 16 kbps 3.8%, MP3 16 kbps 5.4% — twenty-two times smaller for three percentage points. Opus scored better than MP3 at the same bitrate but its one error deleted a person's name, so a lower error rate is not automatically a better transcript. Some errors (rollback heard as rollbook) are vocabulary, not bitrate. - [Extract audio from a video on a Mac](https://kekoso.app/blog/extract-audio-from-video-mac/): three ways
| – |
| 7 Nov 2026 |