FluidVoice turns speech into text on your computer; FluidMeet records, transcribes and labels speakers in meetings. In this The Next New Thing interview, Andrew Warner and FluidVoice creator Prachi Modi show both workflows, then unpack the speech and cleanup models. The app is a compelling free, local-first option, but “local” needs a settings check: optional cloud providers, analytics, screen context and agent access have different privacy implications.
For a Mac, start with plain dictation and one local speech model. Add FluidMeet only if you need a recording and speaker-labeled transcript. Add cleanup, screen context, summaries or agent access one at a time, after checking what each feature can read and where it sends data. The app is open source; its optional Fluid Intelligence runtime is not.
Watch Andrew Warner and Prachi Modi
Andrew demonstrates his own setup and Prachi explains the product. We have not installed FluidVoice, timed its transcription or independently tested its accuracy for this article.
Source Note: The Video and the November Release
We used the supplied transcript and checked the official repository, project site, stable v1.6.12 notes, v1.6.13 beta notes and Mini model research on 4 November 2026. The video calls Fluid Speech Pro unreleased; the official beta now lists it. The stable Mac release includes FluidMeet exports, summaries and agent integrations, while beta features should not be assumed available in stable builds. The project explicitly warns that fluidvoice.org is not an official download site; use the repository releases or the linked official site.
Dictation and Meeting Notes Are Different Jobs
At 03:08, Andrew holds a hotkey and speaks into an empty text field. A live overlay shows recognized words before they land in the app. This is the everyday dictation path: microphone, local speech recognition, optional cleanup, then insertion into the focused field. FluidVoice's quick start explains why it needs both microphone and Accessibility permission for that final step.
FluidMeet solves a different problem. Andrew starts a meeting, then inspects the saved audio, transcript and separate speaker labels. The developer's technical explanation says the Mac meeting pipeline pairs a speech engine such as Parakeet with NVIDIA's Nemotron 3 diarization model to determine who spoke when. The stable release can export timestamped transcripts as Markdown, text or JSON, edit transcript text and speaker names, and connect coding agents from FluidMeet settings.
That last connection matters: Andrew gives an agent instructions to read his meeting library, then asks what he said in the latest meeting. The recording can remain on his Mac while a separately configured agent gains access to the notes. “The app processed audio locally” does not by itself prove every downstream agent or sync service will keep those notes local.
Which Model Does What?
| Component | Job | Practical note |
|---|---|---|
| Fluid Speech Mini | On-device speech recognition for dictation | 253 MB Core ML package. The team's published 4.54% average English word-error figure is its own test, not our benchmark. Check the installed model picker for current language support. |
| Fluid Speech Pico and Pro | Smaller and larger speech-recognition choices | The demo shows roughly 150 MB and 622 MB respectively. Pro was a preview in the video; beta notes now list Pro, Mini and Pico for 25 languages. Sizes and availability may differ by build. |
| Fluid 1.1 Splash / Fluid-1 family | Clean up rough text, punctuation and formatting after recognition | Separate from speech recognition. The video shows Splash using on-screen words to correct names; in the beta, screen capture is optional and off until configured. Check the model download and hardware recommendation in your version. |
| FluidMeet speech + diarization | Transcribe a meeting and label speakers | The documented Mac path uses a chosen speech engine and Nemotron 3 Diarization. A speaker label is an estimate; review names and action items before sharing minutes. |
Open source is not one blanket label for every component. The application is GPLv3; the repo describes Fluid Intelligence as a separate, privately maintained local runtime. Some published speech-model weights have their own license and attribution requirements. Open weights allow use and inspection under those terms, but do not automatically make the training process or every runtime open source.
Get Started on a Mac, Then Test One Meeting
- Use an official download. Install the Mac app from GitHub Releases, or use the repo's Homebrew command
brew install --cask fluidvoice. The README lists macOS 15 or newer; Apple Silicon supports accelerated models, while Intel Macs use the Whisper route. - Grant only the needed permissions. Microphone is needed to capture speech; Accessibility lets the app type into another app. Choose a hotkey and a local speech model in onboarding. Start in a throwaway text editor and check both recognition and insertion before using a work account.
- Keep the first test simple. Dictate a short paragraph with names, punctuation and a correction. Compare raw transcription with optional cleanup. If exact wording matters, retain the raw text and review the rewritten version.
- Try FluidMeet separately. Confirm recording permission with participants, select the intended microphone and meeting source, then record a short low-sensitivity call. Review the transcript, speaker labels, saved audio, export and deletion path. The stable release notes document these meeting controls.
- Only then connect an agent. FluidMeet's Integrations screen exposes setup for Claude Code, Codex and other agents. Inspect the generated skill and the scope of access before handing over a meeting library. Begin with a non-sensitive test transcript rather than the full archive.
The official site lists Windows and iPhone as beta platforms, but the video demonstrates a Mac workflow. The host's Windows early-access link currently resolves to a general site page, not a verifiable FluidMeet sign-up. Use the official releases and check platform-specific feature notes before planning a team rollout.
What Stays Local, and What Can Leave?
In the configured local path, recognition and FluidMeet's recording, transcript and speaker timeline run on the Mac; the team documents that in its FluidMeet post. But the repository's privacy section says anonymous activity signals are uploaded in a weekly batch, with more detailed analytics enabled by default and a setting to reduce them. It says voice and transcript content are not part of those analytics. Optional cloud enhancement or summaries can transmit task content to the selected provider; an agent with access can also read or copy the notes. Screen context is an optional permission in the beta, not a reason to assume the entire screen is always captured.
For genuinely sensitive work, check four boundaries before calling the workflow private: audio storage, chosen AI provider, screen-reading permission and agent integration. Also inspect retention and backups, and follow your organization's recording and consent policy. Local inference reduces one data transfer; it does not replace access control or human review.
What Should You Try First?
| Your bottleneck | Start with | Measure |
|---|---|---|
| Typing interrupts your work | Local dictation, no cleanup | Time to usable text and recognition errors for your accent, language and microphone. |
| Spoken drafts need too much editing | A local cleanup model | Whether it saves edits without changing names, facts or intent. |
| Meeting notes miss who said what | One consented FluidMeet test | Speaker errors, missing decisions and time spent correcting the transcript. |
| An agent needs meeting memory | A single test transcript | Whether retrieved answers are cited correctly and whether the access scope is acceptable. |
The better question is not whether the demo looks instant. It is whether the complete path, including model download, microphone, cleanup, correction and sharing, is faster and more trustworthy on your machine.
Business Pilot: Test Meeting Notes Without Over-Connecting
This original JQ AI Systems prompt frames a small, human-reviewed team pilot. It was written for this article, not supplied by Andrew or Prachi, and has not been tested with a client.
Evaluate one meeting workflow
Consent, accuracy, access and a stop rule.
Sources and Downloads by Role
| Role | Link | Why it matters |
|---|---|---|
| Video demonstration | Andrew Warner with Prachi Modi | Shows Andrew's setup, the live meeting, dictation and agent query; not an independent benchmark. |
| App and downloads | Official repo · Releases · official site | Install routes, license, supported hardware and privacy disclosures. Avoid lookalike download domains. |
| Meeting implementation | FluidMeet technical note · NVIDIA model card | Explains transcription versus speaker labeling and the model's eight-speaker limit. |
| Model and availability | Mini research · published weights · latest beta notes | The team's own test methodology, model-license distinctions and Pro/Splash beta status. |
Video Chapters
The live meeting demo
| Time | Segment |
|---|---|
| 00:00 | Meet FluidVoice |
| 00:20 | Live transcription |
| 00:55 | FluidMeet and speaker identification |
| 01:35 | Audio export and integrations |
| 02:21 | Give an agent access to meeting notes |
Dictation and the model pipeline
| Time | Segment |
|---|---|
| 03:08 | Real-time dictation |
| 03:54 | The models behind it |
| 04:17 | Why local processing matters |
| 04:55 | Meeting models versus dictation models |
| 05:11 | Fluid Speech Mini and Pico |
| 06:04 | Fluid Speech Pro preview |
| 06:47 | Size and accuracy claims |