Free · Open source · Runs on your Mac

Talk, and it types.Speech stays on your Mac.

ThunderTalk is a voice input app for macOS. Press a key in any app, speak, and your words appear at the cursor. Recognition runs on your own machine: no account, no subscription. Cloud engines are optional and use your own key.

Free and MIT licensed. Apple Silicon Mac, macOS 15 or later.

New in v1.14.0 Optional cloud engines with your own key that transcribe while you speak. AI agents can use every feature through the CLI and MCP.
An interactive replica of the Home screen with sample data. Click the key to try a dictation.

01Dictation

One key, any app.

ThunderTalk lives in the menu bar. Wherever your cursor is, whether that is a browser, an editor, Slack or a terminal, the text goes there.

01

Press

Tap your hotkey to start and tap again to stop. Right ⌘ is the default. Prefer to hold it while you talk? Switch the press mode to Hold in Settings.

Right ⌘
02

Speak

A speech model on your Mac's own chip does the listening. Chinese, English, or both in the same sentence. Live Preview shows provisional words in the recording bar while you speak. It is on by default in Settings; Direct translation mode hides it. Preview may pause or stop if the model cannot keep up. Optional background-audio muting never blocks dictation, works on the first press and restores the exact previous volume, even after a crash.

03

Done

Long dictations are never lost: models with a clip-length limit recognize long takes in pieces at your pauses. If recognition still returns empty, Live Preview text is used instead. Every take is saved to History with its recording. After you stop, recognition produces the final pasted text; preview words can change. English terms recognized correctly in a clean, aligned preview can be kept in the final text, and repetition loops are caught and re-decoded. The text goes to your cursor and private, searchable history. Spoken numbers turn into digits, and hotwords stay spelled your way.

Hotwords

Teach it the words it keeps getting wrong: product names, acronyms, colleagues. Available with Qwen3-ASR and Fun-ASR-Nano.

Numbers and AI Proofread

Inverse text normalisation turns spoken numbers into digits in English and Chinese, and leaves titles such as 《一千零一夜》 alone. Optional AI Proofread recovers what you meant with minimal edits to misrecognized words, including real-word homophones, names and jargon. It never translates, rewrites, reformats or changes your language or style.

Open AI Proofread in the sidebar to choose an existing logged-in Codex, Claude Code, Gemini, Grok or Cursor CLI, an installed Ollama / LM Studio model, Cherry Studio's API server or a custom OpenAI-compatible API. Providers are detected automatically with their status; pick a model and a supported reasoning level. No AI model is bundled or downloaded. It applies to plain dictation, not translation. Raw text is pasted first; proofreading replaces it only if you have not typed, clicked, scrolled or switched apps. Temporary failures retry once; errors keep raw text and explain why. Terminal replacement is skipped. After replacement, the overlay shows your full text with old words struck out beside corrections, three lines per page. Hover to pause or click to dismiss.

Speak one language, paste another

Direct mode pastes a translation after you stop speaking; translation to Chinese (Mandarin) now returns Chinese instead of leaving English speech untranslated. Review mode transcribes first, then lets you Replace it or Keep the original. Over 100 languages through SeamlessM4T v2, which needs about 9 GB of disk and is comfortable on 16 GB of RAM or more.

Downloads that tell the truth

Progress comes from the real byte count, an interrupted download resumes, and Cancel stops right away.

A guided first run

Four short steps: welcome, permissions, a model, and a box to try it in.

History on your disk

Every dictation is saved to a file in ~/.thundertalk. Search it from Home, delete a line, or clear it all.

Updates itself

When a release is published a small prompt appears. One click downloads it, swaps the app and relaunches. A stable signature keeps Microphone and Accessibility permissions across future updates; upgrading from v1.7.0 or earlier may ask for them one last time.

02Studio

Compose in v1.11.0

Studio turns recordings into text, text into speech, and lyrics into songs.

Transcribe recordings or web links, save a batch of transcripts, add subtitles, turn text into speech, or compose an original song. Audio processing stays on your Mac. Optional AI meeting notes use your chosen provider.

Transcribe

Drop in a recording or paste a link, get a timestamped transcript.

Audio or video: m4a, mp3, wav, mp4, mov and more. Choose a downloaded speech model for each job; your dictation model is selected by default. For speaker labels, choose MOSS-Transcribe-Diarize and enable Label speakers; click a speaker to rename it. Export TXT, Markdown, SRT, VTT or JSON. The page stays compact after transcription, and timestamps and plain text switch instantly even for hour-long recordings. Common formats decode with macOS tools; WebM, MKV and some other formats may need ffmpeg.

Paste single-video links from YouTube, Bilibili or other yt-dlp-supported sites; yt-dlp ships with the app. Audio downloads temporarily, is transcribed locally, then is deleted. Members-only and sign-in restrictions are flagged; the app does not sign in for you. A noticeably shorter preview than the advertised duration is flagged as a partial transcript. Playlists, channels and live streams are unsupported.

Add several files or links to the queue, choose formats and a destination folder, then Transcribe all. Items run one at a time and save automatically. Or save beside source files, with link results in Downloads. Cancel one item or the whole queue. Successful items leave the queue automatically; failed or cancelled items stay for retry or removal. Every completed transcript is saved in History: reopen, rename, delete or search it by title or text. Duplicate export names get a numbered suffix.

After transcribing a local video, choose Add subtitles to video… to burn always-visible subtitles into a copy or add a subtitle track the viewer can turn on or off. It needs ffmpeg: with Homebrew installed, run brew install ffmpeg. Link transcription downloads audio only, so burn-in needs a local video.

AI meeting notes turn a transcript into a summary, key points, decisions, action items and open questions in English or Chinese. Reuse the provider/model selected on the AI Proofread sidebar page, even with proofreading off; no AI model is downloaded. Copy or save Markdown, include notes in a Markdown transcript export, or enable notes after each queue item. Review notes against the transcript; failed notes keep the transcript.

Interactive Transcribe preview
onboarding-review.m4a
click to choose another
Sample
Export

Click a speaker name to rename it. The exports update as you type.

Speak

System voices, or your own voice.

Four local engines: the Chinese and English voices installed on your Mac, including Siri voices, appear with the System tag. They need no download, render about 37× faster than real time in our tests, and do not clone voices. VoxCPM2 offers studio-quality 48 kHz speech, IndexTTS-2.5 faithful cloning, and Kokoro small and fast speech; these three download once when you first pick one of their voices. Set the speed, scrub through the result, and save it as WAV or M4A.

Interactive Speak preview. This page plays no sound.
Built-in voices
My voicesClone engine

Built-in VoxCPM2 and IndexTTS voices are designed from a text description, not recorded from real people.

Speed
Language: detect automatically
0:00 / 0:07
Clara (EN) · VoxCPM2 · 1.0×

Preview only: no audio is played and no file is written.

My voices

Clone your voice from 5 to 15 seconds of audio.

Record yourself reading a short passage, or import a clean recording. The reference is kept on your Mac under My voices, ready whenever you want to hear your own voice read something. It is spoken with VoxCPM2 or IndexTTS, whichever you choose. Measured on an M3 Max, cloning a 3.6-second recording of the maintainer: an independent speech recogniser read the result back with 0 to 3.4% error, and speaker similarity was 0.97 to 0.98, with either engine. Only clone voices you have the right to use.

  1. 01
    Record or importFive to fifteen seconds of natural speech in a quiet room, or a clean recording of one person.
  2. 02
    Check the wordsThe recording is written out on your Mac. A wrong transcript makes the clone worse, so correct it.
  3. 03
    Name it and saveIt appears under My voices, next to the built-in ones.

After downloads, file transcription and speech generation work offline. Web links and cloud AI providers need a connection.

Simulated voice recording. This page never uses your microphone.
Record 0 s

Simulated. This page never touches your microphone.
My voices
Nothing saved yet

Compose · ACE-Step 1.5 · MIT

Original songs on your Mac.

Create vocals with background music or an instrumental. Turbo is the fast default; optional XL gives clearer lyric read-back and needs at least 32 GB RAM plus plenty of free memory. Compose is hidden below 16 GB; 16–23 GB is experimental turbo only, one take, up to four minutes. The memory guard refuses jobs when headroom is short, and dictation keeps priority.

Qwen3-ASR flags lines that were not sung clearly; listen to confirm. Adjust difficult words with the sung-lyrics pronunciation helper, review the approximate range to repaint a line when memory allows, or make a new take. Repaint cannot fit the 16 GB tier. Takes save to ~/Music/ThunderTalk. Describe style by genre, mood and instruments; real artist names trigger a warning. When releasing to streaming platforms, disclose AI-generated music through your platform or distributor.

03Privacy

Your voice has nowhere to go.

Recognition, translation, transcription and speech synthesis run on your own Mac; audio is never uploaded. Optional AI Proofread and meeting notes can send text to the provider you choose. The code is open, so you can check how it works.

Audio
Recognised on your Mac and never sent anywhere.
History
A plain file in ~/.thundertalk on your disk. Search it, delete a line, or clear it.
Network
Model/component downloads, GitHub update checks and Studio links. Optional proofreading/notes send text to your chosen provider: cloud CLIs use your account; Ollama and LM Studio use local servers; Cherry Studio and custom APIs may forward text to cloud models. AI Proofread is off by default. Optional cloud speech engines (off until you choose one) send audio or text to that provider with your own key, kept in the macOS Keychain.
Account
No ThunderTalk account, subscription or usage limit. Optional AI providers use your own account or server.
Code
MIT licensed. Read it, build it yourself, fork it.
Speech recognition turns your audio into text at the cursor on your Mac. Audio is not uploaded; optional AI Proofread can send text to your chosen provider. Internet On your Mac Audio stays on your Mac Your voice microphone Speech model runs on your chip Text at the cursor any app Audio in, text out. No network hop.

04Models

Choose the engine that suits your Mac.

Every speech model runs on-device. Qwen3-ASR-0.6B ONNX stays recommended on every Mac. Apple · Built in needs no download: about 3× faster than Qwen3-ASR ONNX in our tests and fine for clear, close speech, but weak on noisy, far-away, fast or sung audio and mixed Chinese-English speech. macOS asks for Speech Recognition access once when you activate it. The app reads your hardware and marks the models that fit; downloads show real progress. The Models page shows parameter counts, download sizes, CPU or Apple GPU requirements, language tags (including an explicit “No Chinese” where applicable) and features. Only the dictation model stays resident; Studio, text-to-speech and translation models release after three idle minutes and reload when needed. Idle memory fell from about 9 GB to under 2 GB. AI Proofread and notes use your existing provider/model selected on the AI Proofread sidebar page; no AI model is bundled or downloaded for them.

The v1.7.0 download measured about 592 MiB installed and 224 MB zipped (214 MiB), down from 961 MiB and a 330 MB zip in v1.6.4. Translation and IndexTTS share an optional 90 MB PyTorch component, downloaded by the app when needed into ~/.thundertalk/runtime/. Restart once if prompted. Dictation, Studio transcription, VoxCPM2 and Kokoro do not need it. Model weights below are extra.

CLI v2: ThunderTalk --cli transcribe|speak|voices|status runs without a window, even while the app is open. Progress and results are JSON lines. Transcribe exports TXT/Markdown/SRT/VTT/JSON; --engine best selects the most accurate downloaded ASR model, and --speakers uses MOSS for speaker labels. Speak exports M4A/WAV/MP3; its --engine best prefers IndexTTS, then VoxCPM2, Kokoro (except mixed Chinese-English text), then Apple. Packaged apps advertise ThunderTalkCLIVersion 2. The MCP server can call it.

ModelRuns onDownload sizeLanguagesGood for
Speech recognition
Apple · Built inCPU / Neural EngineNo downloadDictation languages installed on your MacOn-device only. Quick start or fallback for clear speech; Qwen3-ASR stays recommended.
SenseVoice-SmallCPU · ONNX241 MB5 languagesSmallest and quickest to start. No hotwords.
Qwen3-ASR 0.6B int8 RecommendedCPU · ONNX940 MB52 languagesRecommended on every Mac. Supports hotwords.
Qwen3-ASR 0.6BGPU · MLX~1.9 GB52 languagesApple GPU option. Supports hotwords.
Qwen3-ASR 1.7BGPU · MLX~4.7 GB52 languagesHarder accents and noisy audio. Wants 16 GB of RAM.
MOSS-Transcribe-Diarize 0.9B StudioGPU · MLX~1.8 GB50+ languagesSpeaker labels and timestamps. Powers the multi-speaker mode.
Parakeet-TDT 0.6B v3CPU · ONNX640 MB25 European languages · No ChinesePunctuation and casing built in.
Parakeet-TDT 0.6B v2CPU · ONNX640 MBEnglish · No ChineseFast English dictation on the CPU (about 50× real time on an M3 Max).
Fun-ASR-Nano-2512 0.8B ExperimentalCPU · ONNX842 MBChinese, English, JapaneseSupports hotwords. In our tests, roughly 25% faster on short sentences than Qwen3-ASR-0.6B, with similar accuracy and lower memory. About 20 s per pass; longer dictations are split at pauses. Can invent text on pure noise.
Studio
Apple system voicesCPU · AVSpeechSynthesizerNo downloadChinese, EnglishInstalled voices, including Siri, tagged System in Speak. About 37× real time in our tests. No cloning.
VoxCPM2 8-bit SpeakGPU · MLX3.2 GBChinese, English and moreStudio-quality 48 kHz speech and voice cloning, about real time. Apache-2.0.
IndexTTS-2.5 8-bit SpeakGPU · MLX1.7 + 2.3 GB5 languagesVery faithful voice cloning in Chinese, English, Japanese, Spanish and Arabic, about real time. 1.7 GB model plus a 2.3 GB encoder. bilibili Model Use License.
Kokoro v1.1 82M SpeakCPU · ONNX364 MBChinese, EnglishSmall and fast, about 4× real time. 23 built-in voices (20 Chinese, 3 English). Apache-2.0.
Music · Compose
ACE-Step 1.5 Turbo 2BGPU · MLX / MPS~6.3 GBSongs / instrumentalsMIT. Default and recommended; ≥16 GB RAM. Experimental at 16–23 GB, one take, ≤4 min.
ACE-Step 1.5 XL 4BGPU · MLX / MPS~21.5 GBSongs / instrumentalsMIT. Optional, clearer lyric read-back; ≥32 GB RAM and substantial free memory.
Translation
SeamlessM4T v2 LargePyTorch · MPS / CPU~9 GB100+ languagesSpeech and text translation. Best with 16 GB of RAM or more.

One-time catalog downloads above include shared VAE and text-encoder weights. XL's checkpoint alone is about 20 GB and is converted to bf16 after download. The optional 1.7B planner adds about 3.8 GB (turbo total ~10.1 GB; XL ~25.3 GB). Shared weights are reused. Runtime components and lyric-check ASR weights are separate. Music generation works offline after downloads.

05Compare

How it compares.

Taken from each product's own website on 29 September 2026. Prices change, so follow the links.

ThunderTalkWispr FlowSuperwhisperVoiceInk
PriceFreeFree up to 2,000 words a week; Pro $15 a month ($12 billed yearly)Free plan; Pro subscription or lifetime licenceFree trial; paid licence
Open sourceYes, MITNoNoYes, GPL-3.0
Where speech is processedOn your Mac (cloud optional, your own key)In the cloud (their security FAQ)On your device or in the cloud, your choiceOn your Mac
PlatformsmacOS 15 or later (Apple Silicon)Mac, Windows, iOS, AndroidMac, Windows, iOS, AndroidmacOS 15 or later

ThunderTalk also transcribes meetings with speaker labels, reads text aloud with four local speech engines and clones your own voice, all on your Mac. We have not checked whether the others offer these, so they are not in the table. ThunderTalk needs an Apple Silicon Mac with macOS 15 or later. It is signed with an Apple Developer ID and notarised by Apple.

06Questions

Plain answers.

Something not covered here? Open an issue on GitHub and it will get an answer.

Is ThunderTalk free?

Yes. ThunderTalk is MIT licensed and free to use. There is no account, no subscription and no usage limit, and the source code is on GitHub.

Does it work offline?

Yes. After downloading the models and any required component, speech recognition, translation, file transcription and text to speech work offline. Studio links need a connection. Optional AI Proofread and meeting notes depend on your chosen provider: Ollama and LM Studio use local servers; cloud providers need a connection and receive text. Optional cloud speech engines need a connection.

Which apps can I dictate into?

Any app with a text cursor: browsers, editors, Slack, mail, terminals. ThunderTalk pastes the text where the cursor is, so it asks for Microphone and Accessibility permission the first time you run it.

What is Studio?

Studio has Transcribe, Speak and Compose. Compose generates original songs locally with ACE-Step 1.5: turbo by default, optional XL on Macs with at least 32 GB RAM and enough free memory. Transcribe handles audio/video files and single-video links from YouTube, Bilibili and other yt-dlp-supported sites, with a model choice per job, timestamps, optional speaker labels, batch exports and subtitle burn-in for local videos. Every completed transcript goes to History. The layout stays compact, with instant timestamp/text switching. Export TXT, Markdown, SRT, VTT or JSON. AI meeting notes reuse the provider/model on the AI Proofread sidebar page, even with proofreading off. Notes support up to 16 parts of about 6,000 characters plus one merge, a 60-second timeout per call and Cancel; split oversized recordings. Speak uses local VoxCPM2, IndexTTS-2.5 or Kokoro, with built-in voices or a voice cloned from a 5 to 15 second recording. Links and cloud AI providers need a connection.

Do I need ffmpeg for Studio?

Common formats such as M4A, MP3, WAV, MP4 and MOV decode with macOS tools. Burning subtitles into a video copy requires ffmpeg, and WebM/Opus link audio or other formats may need it too. ffmpeg is not bundled. With Homebrew installed, run brew install ffmpeg in Terminal, then retry.

Where does a cloned voice go?

Nowhere. The short reference recording and its transcript are stored in ~/.thundertalk/voices on your Mac, and nothing is uploaded. Delete a voice in the app, or delete the folder, to remove it. Please only clone voices you have the right to use.

Does it handle Chinese and mixed Chinese-English?

Yes. Qwen3-ASR handles 52 languages, including Chinese and English mixed in the same sentence, and the app itself is available in English and Chinese.

Which Macs are supported?

An Apple Silicon Mac (M1 or newer) with macOS 15 or later. The download is built for Apple Silicon only, and the bundled speech libraries need macOS 15. Intel Macs are not supported by the release build.

Will macOS warn me when I open it?

No. From v1.9.2, ThunderTalk is signed with an Apple Developer ID and notarised by Apple, so it opens like any other downloaded app. Updating from v1.9.1 or earlier asks for Microphone and Accessibility once more; after that, permissions stay granted across updates.

How is it different from Typeless, Wispr Flow or superwhisper?

They are closed-source products with paid plans; Wispr Flow processes speech in the cloud, and Superwhisper lets you choose local or cloud models. ThunderTalk is free and open source, with speech processed on your Mac. Optional AI Proofread and notes use your selected provider. See the comparison for sources.