Skip to content
OxSay app icon: a glass microphone beside three lines of text on a blue-to-violet square.

Dictate, record calls, voice a script.All on your Mac.

Local, private, with an MCP server for agents.

Free while in beta. macOS 14.2 or later. Apple Silicon, or Intel with AVX2.

OxSay's Meetings window. On the left, recorded calls listed by day, with a 42-minute Zoom call selected. On the right, that call's notes: the four people on the call, a three-sentence summary, two decisions, and action items with an owner and a date.
Sample data.

Hold ⌃Space. Speak.

In any app. Let go, and the text appears at the cursor.

OxSay's Dictation window. The shortcut, Control-Space, above the line “Press the key, speak, and your words appear where the cursor is.” Below, the choice between holding the key and pressing it once, and four recent dictations marked “Kept in memory only.”
Sample data.

Text in about a second

A short dictation turns into text about a second after you let go: 0.9 s measured when the unfinished tail was short.1

It quiets the room and minds the clipboard

While you dictate, the app can pause Apple Music and Spotify or lower everything else the Mac plays. Your previous clipboard comes back afterwards, and a focused password field gets no paste.

25 European languages

Including Russian and English, with NVIDIA's Parakeet TDT 0.6B v3. Whisper large-v3-turbo is an optional second recognizer with 99 languages.

Calls, recorded and written up.

OxSay records a call, writes the transcript, tells the other people's voices apart and writes notes.

“Record this call?”

The prompt appears when a call starts in Zoom, Teams, Google Meet (in Chrome, Brave, Arc or Edge), Webex, Slack, FaceTime, Discord, WhatsApp or Telegram.

Per app, you choose to be asked, to record automatically, or to stay off. A red REC badge stays on screen during every recording.

Sample data.

OxSay's Settings, the list “Calls in these apps”: Zoom, Google Meet, Microsoft Teams, Webex, OxPulse, Slack, FaceTime, Discord, WhatsApp and Telegram, each with its own menu set to “As set above (Ask me first)”.

Who said what

The two channels are merged into one conversation. On Apple Silicon with macOS 15 or later, the other side's voices are told apart; you can name a speaker, and the transcript and notes are rewritten with the names.

Click any line of the transcript to play the audio from there.

Sample data.

The Transcript tab of a recorded Zoom call. Four turns from You, Anna and Boris, each with a name and a timestamp; Boris's line is highlighted and marked with a speaker symbol because its audio is playing.

Notes, in the language of the meeting

A local language model writes a summary, decisions, action items, open questions and key numbers.

An action item names the person who owns it only when the transcript shows it, and quotes the words that show it.

Sample data.

The Notes tab of the same call: the people on the call, a summary, two decisions and three action items, each with an owner and a date.

Find things

Search covers titles, notes and transcripts.

Your microphone and the Mac's sound

Two channels that start together: your microphone and everything the Mac plays.

Notes on a 16 GB Mac

With 16 to 23 GB the notes model is Qwen3.5-4B (2.7 GB). It never starts below its floor, so it does not squeeze the browser that runs your call.

Recording other people is your responsibility: the law in some places requires everyone's consent. The app can copy a sentence for that to the clipboard.

Files in, text and subtitles out.

Drop an audio or video file on the Transcribe page, or pick one.

From a recording to .txt and .srt

WAV, MP3, M4A, AIFF, FLAC, MP4, MOV and similar files work, up to 200 MB by default. You get the text and subtitles; .vtt and .json are under More.

A long file keeps working if you close the window. The audio you drop is not kept.

Time to transcribe 6 minutes of audio on an Apple M5 Pro: about 2.5 s.2

Sample data.

OxSay's Transcribe window: a drop area labelled “Drop a recording here”, one file marked “Transcribing…”, and a list of finished files, each with Text, Subtitles and More buttons.

Clean up the voice in a video

Give OxSay a video or audio file and get a copy with the voice cleaned of noise, room sound and other sounds under it.

The video, other audio streams and chapters are copied unchanged. It needs macOS 15 or later.

It is a file tool, reached through the MCP tool voice_clean and the HTTP API (POST /v1/audio/clean), not the app window.

A script in. A voiceover out.

Write or paste a script, or open a text, Markdown, Word, RTF, OpenDocument or PDF file. Pick a voice and press Create Voiceover.

Standard, or Studio

Standard is the fast voice: it suits short notices, and first audio arrives in a fraction of a second.

Studio is the expressive voice, for Reels, YouTube and podcasts. It needs Apple Silicon with 24 GB or more and an opt-in download of about 3 GB.

A voiceover can be stopped, resumed after an interruption or a closed laptop, and paused on battery power.

Sample data.

OxSay's Voiceover window: a two-sentence script, the line “A short video for Reels and Shorts: shorter pauses, louder sound”, a voice menu and a quality switch set to Studio.

Change one phrase, not the take

A finished voiceover is listed with its phrases. Each phrase can get a mood and another variant, and the change touches that phrase only.

The New reel preset is tuned for short vertical video.

Sample data.

A finished voiceover named “Reel: the first one”: its waveform, a 41-second player, and four phrases, each with a mood menu (Warm, As usual) and a variant count.

The command line and the API also write subtitles (.srt, .vtt) and the time of every word for animated captions.

Your voice. Or one you describe.

The Voices page holds your voices and the voices used to read selected text aloud.

Clone a voice in a few seconds

Record yourself in the app: read the shown text for a few seconds. Or add a voice from an audio file.

Sample data.

The “Record your own voice” sheet on the Voices page: a paragraph to read aloud, a red recording dot at 7.5 seconds with moving level bars, the line “OxSay listens only while you are recording”, and a Stop button.

Describe a voice in words

“A warm, deep male narrator in his forties.” Hear several versions and keep the one you like. This uses the studio voice.

Sample data.

The “Describe a voice” sheet: the description “A calm male narrator, unhurried and warm”, a sample sentence in Russian, the name “My narrator”, and Cancel and Create buttons.

Read aloud, anywhere

Select text in any app and press ⌃⌥Space, and OxSay speaks it in a voice you pick per language.

A voice per language

Pick the voice for Russian and English text separately, or let the app choose a random one each time.

What stays on the Mac, what leaves it.

Recognition, notes, voices and cleanup run in local processes.

Stays on the Mac

  • Meetings, transcripts, voices, models and your glossary live under ~/Library/Application Support/ox-say.
  • Dictated text is held in memory and never written to disk.
  • Meeting audio is deleted after 7, 30 (default) or 90 days, or never; transcripts and notes stay until you delete the meeting.
  • The daemon listens on 127.0.0.1 only and refuses requests a web page could send to it.

Leaves the Mac

  • The only network traffic is to huggingface.co for model files and to github.com for the installer, release packages and the update check.
  • The repository says the hosts in the code are these two and the loopback address.

Two things to know

A copy of each meeting's notes goes to a notes folder you choose, ~/Documents/ox-say Meetings by default, and macOS uploads that folder if “Desktop & Documents” iCloud sync is on; the app warns you in Settings.

The uninstaller's --purge leaves meetings, transcripts and the glossary in place, so you delete those folders yourself.

An MCP server for your agents.

The installer registers it with Claude Code. For another client:

claude mcp add --transport http --scope user ox-say http://127.0.0.1:8094/mcp

The tools

speak, transcribe, voice management and design, long-form produce, voice_clean, and read-only meeting access (meetings_list, meeting_get, meetings_search).

An OpenAI-compatible API

The daemon on 127.0.0.1:8094 also serves /v1/audio/speech and /v1/audio/transcriptions.

Meeting text is data

Meeting texts quote other people's speech, so a client should treat them as data, never as instructions.

Coming next, and not built yet: live captions while a meeting is recording (#527) and a live meeting transcript for your own agents (#549).

How it gets names right.

Speech recognizers fumble names. OxSay keeps a glossary of corrections, each one a pair.

ВоксейWhat the recognizer wrote
OxSayWhat you meant

Replaced after every recognition

In dictation, file transcripts, meeting transcripts and the MCP transcribe tool. Matching is whole-word, ignores case and treats ё as е.

Nothing enters the glossary silently

An agent can file a pair. It stays a candidate until you approve it, or until it has been filed from enough different sentences. A rejected pair stays rejected.

Spoken-form boosting passes the spoken forms in your glossary to the decoder as a hint, so it prefers them and the correct written name comes back. Treat it as a help for names you correct often, not as a general accuracy gain.3

Runs on Intel Macs too.

The modern local speech stacks have left Intel Macs behind. OxSay carries the fixes.

Current speech models on a 2019 MacBook Pro's GPU

A 2019 to 2020 MacBook Pro with a Radeon Pro 5300M, 5500M or 5600M is the tested case.

60 s of audio decoded in 2.50 s on Metal (0.38 s on an M5 Pro).4

The 0.6B fast voice runs faster than real time: RTF 0.58 to 0.61.4

Works on Intel

Dictation, Transcribe, the fast voice and the meeting transcript.

Needs Apple Silicon

Telling the voices of a call apart, and the studio voice. On Intel the notes model runs on the CPU, and a notes run can take many minutes.

PyTorch stopped publishing macOS x86_64 wheels after 2.2, MLX runs only on Apple Silicon, and stock ggml gives wrong results on discrete, non-unified-memory Metal GPUs.

Install and requirements.

One line in Terminal. It never uses sudo.

curl -fsSL https://raw.githubusercontent.com/anatolykoptev/ox-say/main/get.sh | sh

The installer downloads the release for your Mac, checks its SHA-256, fetches checksummed models, starts the daemon as a LaunchAgent, runs a speak-and-transcribe self-test and installs the app. Prefer a package? Download from GitHub releases.

macOS 14.2 or later

Separating the voices of a call and voice cleanup need macOS 15 or later.

Apple Silicon, or Intel with AVX2

On an Intel Mac the far side of a call is one speaker, notes run on the CPU and can take many minutes, and the studio voice is not available.

About 2.5 GB of disk

At install on Apple Silicon; about 2.3 GB on Intel. The notes model (2.7 to 5.7 GB) downloads the first time you ask for notes.

Memory on Apple Silicon

Under 16 GB gets transcripts only, no notes; 16 to 23 GB gets the 4B notes model; 24 GB or more gets the 9B notes model and the studio voice.

Before first use

Needs ffmpeg from Homebrew. First launch asks for the microphone and Accessibility, which lets the app paste into other apps. The studio voice (about 3 GB) and Whisper (1.6 GB) are opt-in.

Time to transcribe 6 minutes of audio on an Apple M5 Pro: about 2.5 s.5

OxSay app icon: a glass microphone beside three lines of text on a blue-to-violet square.

Get OxSay.

Free while in beta.

macOS packages for arm64 and x86_64 with a SHA256SUMS file.

1. Source: ox-say docs/dictation.md, as cited in the README on origin/main, read 2026-11-15.

2. Source: ox-say docs/engines.md; one measurement, release v0.1.32; as of 2026-11-14. Documented: ox-say docs/engines.md and README on origin/main, read 2026-11-14.

3. On synthetic speech (OxSay's own text-to-speech voices, with the glossary held out by voice), Russian name recall rose 21 to 24 points on 14 terms, with no false insertions across 154 Russian and 60 English clips that did not contain the terms. It has not been measured on real speakers, and a name roster on 20-minute public-meeting recordings moved only 2 of 379 names.

4. Source: ox-say docs/performance.md, Core i9-9880H with Radeon Pro 5500M (4 GB), 2026-10-14; warm whole-clip decode, median of 3 runs. The Metal figure was measured with both engines allowed on the GPU; on a 4 GB card the daemon by default keeps the resident recognizer on the CPU, where the same 60 s clip takes 5.05 s.

OxSay is pre-1.0, with a release most weeks. The app ships in English and Russian. Its code is Apache-2.0; each model keeps its publisher's licence: the default models are Apache-2.0 and CC-BY-4.0, and Parakeet is CC-BY-4.0 from NVIDIA. The studio voice model's card forbids impersonation, fraud and disinformation and asks that AI-generated speech be labelled as such.

5. Source: ox-say docs/engines.md; one measurement, release v0.1.32; as of 2026-11-14. Documented: ox-say docs/engines.md and README on origin/main, read 2026-11-14.