Free, GPLv3, and each with its own tuned AI model.
FluidVoice is the rare case where both sides ship a custom AI model built specifically for cleaning up dictation. This page is about where the two apps genuinely differ: macOS version support, engine choice, and what happens if you cancel by accident.
Quick answer
FluidVoice or EnviousWispr: which should you choose?
Both apps are free, GPLv3, and can transcribe and polish speech on your Mac. Choose FluidVoice if you want fourteen speech-model choices, an Intel or Windows route, file transcription, or deeper prompt routing. Choose EnviousWispr if you want a smaller setup on macOS 14 or later, direct paste with Escape Recovery, and a built-in polish model whose base model and license are published. FluidVoice is the more configurable toolkit; EnviousWispr is the more guided default. Neither choice requires a subscription or account, so the practical decision is flexibility versus simplicity rather than price. If you dictate mainly on a newer Mac, both cover the core job well. The stronger reason to switch is a missing workflow or compatibility feature, not a basic difference in transcription privacy.
Still deciding? See the best Mac dictation apps by use case or browse all 17 dictation app comparisons.
Feature comparison
An honest look at how the two tools stack up across the dimensions that matter most.
| EnviousWispr | FluidVoice | |
|---|---|---|
| Price | Free. No subscription. | Free. No subscription. |
| Account required | No | No |
| Source code | Open source (GPLv3) | Open source (GPLv3) |
| macOS required | 14 Sonoma or later | 15 Sequoia or later |
| Speech engines | Parakeet TDT v3 + WhisperKit, switch in Settings | 14 visible choices across Apple Speech, Cohere, Nemotron, Parakeet, and Whisper |
| AI polish (cleanup after transcription) | EG-1 (our own tuned model), Apple Intelligence, Ollama, or your own OpenAI, Claude, or Gemini key | Fluid-1 plus broad provider choice: OpenAI, Anthropic, xAI, Groq, Cerebras, Google, OpenRouter, Ollama, LM Studio, or a custom endpoint |
| Needs a prompt to be written | No: EG-1 and Apple Intelligence ship with a working default | No: a working default ships with the app, and prompt profiles can be edited or routed by app |
| AI polish model's license disclosed | Yes: EG-1 Community Model License, published, base model named (Qwen3-4B-Instruct-2507) | Not published in the app's own repository as of 2026-08-21 |
| Live preview while speaking | Yes, off by default | Yes, real-time transcription overlay |
| Recover a cancelled dictation | Yes (Escape Recovery, opt-in, keybind/triple-press cancel only) | Not documented |
| Custom vocabulary | Yes (names, terms, jargon; imports from 8 other apps, FluidVoice included) | Yes (aliases, import/export, auto-learning, and Train by Voice) |
| Transcription latency | 0.61s median; ~1.65s with on-device AI polish | Not published |
| Platform support today | macOS (Apple Silicon only) | macOS (Apple Silicon + Intel via Whisper). A Windows pre-build is available; iOS is announced, not yet shipped. |
| Multi-language | 25 European (Parakeet), 99+ via WhisperKit | Up to 99 via Whisper; 40 via Nemotron, 25 via Parakeet, 14 via Cohere, depending on engine chosen |
| Text lands in the right app | Yes (target app reactivation via Accessibility API, three-tier paste fallback) | Yes (direct insertion or clipboard paste, configurable) |
FluidVoice claims sourced from github.com/altic-dev/FluidVoice (README, repo description), the GitHub API, prior internal research (2026-06-23), and FluidVoice v1.6.9 opened directly. EnviousWispr latency from production PostHog data on Apple Silicon Macs. Competitor claims last verified: 2026-08-21.
Both ship a tuned AI model. Here is what else EnviousWispr adds
FluidVoice built something genuinely impressive in Fluid-1, a model tuned for the exact problem we tuned EG-1 for. This is what separates the two apps around that shared idea.
EnviousWispr runs on macOS 14 Sonoma and later. FluidVoice needs macOS 15 Sequoia or later, one version newer. If you have not updated yet, EnviousWispr still works for you.
Escape Recovery keeps a dictation you cancel with your keybind or a triple press, instead of discarding it instantly. Off by default; you choose to turn it on. We found no equivalent documented for FluidVoice.
EG-1 ships under a published EG-1 Community Model License, and we name its base model, a fine-tune of Qwen3-4B-Instruct-2507. FluidVoice's Fluid-1 model and runtime are not part of its GPLv3 app repository and we found no equivalent public license for the model itself.
Measured in production on Apple Silicon Macs. Parakeet TDT runs on the Neural Engine specifically tuned for fast, accurate dictation, so text appears almost instantly after you stop talking.
Built-in filler-word removal runs before AI polish, and it is language-aware: with your language locked to German, Dutch, Danish, or Norwegian in Settings, it keeps real words like "er" and "um" instead of stripping them as English filler.
Smart Import reads your custom vocabulary straight out of FluidVoice's own files, without changing them, and shows you every word before anything joins your library. No retyping.
Where does your voice go?
Both apps process your voice on your Mac by default. For a deeper dive, read on-device vs cloud dictation privacy.
Parakeet TDT: built for speed
EnviousWispr uses NVIDIA Parakeet TDT v3, a multilingual ASR model covering 25 European languages that runs natively on the Apple Neural Engine. No network dependency, no server queues.
Based on production data from Apple Silicon Macs. Results vary by hardware and settings. FluidVoice does not publish comparable latency figures.
Two tuned models, two different disclosures
Both apps solved the same problem: a general-purpose transcript needs cleanup a raw ASR model cannot do on its own. The difference is what each project tells you about how.
FluidVoice 1.6.9 showed fourteen choices across Apple Speech, Cohere, Nemotron, Parakeet, and Whisper. The list includes streaming, multilingual, small, and large options with different speed and language trade-offs.
EnviousWispr keeps it to two: Parakeet TDT v3 for the 25 languages it covers best, and WhisperKit filling in the other 74, switched manually under Settings, Transcription. There is no larger menu; the app is built around this one pairing.
Fluid-1 is a genuine achievement: a model fine-tuned specifically for turning rough dictation into clean text, shipped free with the app, no prompt required. EG-1 exists for the same reason and works the same way. Neither app makes you write your own cleanup instructions.
Where they differ is disclosure. Our README names EG-1's base model, states its license terms, and explains what you can and cannot do with it. FluidVoice's app code is GPLv3, but the Fluid-1 runtime and weights sit outside that repository, and we could not find a published license or model card for it as of this writing. That may simply mean it has not been written yet, not that it does not exist; we welcome a correction if one is published.
Choose FluidVoice if any of this matters more to you
FluidVoice is a genuinely capable, free, open source project with real technical ambition. It may be the better fit here:
FluidVoice 1.6.9 showed fourteen speech-model choices across five families. EnviousWispr offers Parakeet and WhisperKit, which is simpler but leaves much less to tune.
FluidVoice's Whisper option supports Intel Macs. EnviousWispr requires Apple Silicon.
FluidVoice has over 10,000 stars on GitHub as of August 2026, well ahead of our own repository, along with an active following that grew quickly after launch.
FluidVoice's own GitHub repository states a Windows pre-build is already available, with iOS support announced but not yet shipped. EnviousWispr is macOS only, with no other platform in the works.
FluidVoice can transcribe an existing audio or video file you drag in, WAV, MP3, M4A, OGG, MP4, and more, useful for a meeting recording made elsewhere. EnviousWispr only transcribes what you dictate live; it has no file-upload transcription today.
If you want fourteen speech-model choices, deep prompt routing, an Intel Mac path, or file transcription, FluidVoice is a strong pick. If you want a smaller setup with a tuned default model on more macOS versions, give EnviousWispr a try. You can always use both.
Common questions
Yes. Both EnviousWispr and FluidVoice are completely free, with no subscription and no account required. Both are also open source under the same license, GPLv3, so you can read the code that runs on your Mac.
Both are Mac-native, on-device, GPLv3, and ship a custom AI model tuned specifically for cleaning up dictation. EnviousWispr supports macOS 14 Sonoma and later; FluidVoice needs macOS 15 Sequoia or later. FluidVoice 1.6.9 showed 14 speech-model choices across Apple Speech, Cohere, Nemotron, Parakeet, and Whisper; EnviousWispr offers Parakeet TDT and WhisperKit. FluidVoice also has deeper prompt routing, spoken formatting, file transcription, and Command Mode. EnviousWispr keeps setup smaller and adds Escape Recovery for a dictation cancelled by accident.
Yes. FluidVoice ships Fluid Intelligence, also called Fluid-1, and a working default prompt. It also supports OpenAI, Anthropic, xAI, Groq, Cerebras, Google, OpenRouter, Ollama, LM Studio, and custom OpenAI-compatible endpoints, with reusable prompt profiles and app overrides. EnviousWispr offers a smaller provider list and its own EG-1 model with published license and base-model details.
Yes. After the one-time model download, transcription runs entirely on-device with no internet required. AI polish can also run offline using EG-1 (our own model, macOS 14+), Apple Intelligence (macOS 26+), or Ollama with a local model. Cloud providers like OpenAI, Claude, or Gemini are optional and require your own API key.
Yes, and under the exact same license. Both EnviousWispr and FluidVoice's app code are licensed under the GNU General Public License v3 (GPLv3), an OSI-approved copyleft license. In both cases, the polish model itself sits outside that license under its own separate terms: our EG-1 under the EG-1 Community Model License, and FluidVoice's Fluid-1 under terms it has not published in the same place as its app code.
Yes. EnviousWispr's Smart Import reads your custom vocabulary directly out of FluidVoice, without changing FluidVoice's own files, and shows you every word it found before anything joins your library. It also imports from Wispr Flow, Superwhisper, Handy, TypeWhisper, Spokenly, Juno, and Vox.
EnviousWispr needs macOS 14 Sonoma or later. FluidVoice needs macOS 15 Sequoia or later, one version newer. Both need Apple Silicon for their primary engines, though FluidVoice also supports Intel Macs through its Whisper option.
EnviousWispr runs NVIDIA Parakeet TDT v3 for 25 European languages and WhisperKit for 99+ languages, switched manually under Settings, Transcription. FluidVoice 1.6.9 showed 14 choices: two Apple speech options, one Cohere model, three Parakeet models, two Nemotron models, and six Whisper sizes.
Windows, partly. FluidVoice's own GitHub repository description states a Windows pre-build is available, so a Windows version exists though it may be earlier-stage than the Mac app. iOS is not shipped; the repository lists it as planned. EnviousWispr is macOS only, on Apple Silicon specifically.
Yes. Download EnviousWispr, set your keybind, and Smart Import can bring your FluidVoice custom words across in one pass. There is no other data to migrate. Both apps work in any text field on macOS, so nothing about your workflow has to change.
Free, GPLv3, and it runs on more Macs.
Free to download. No account required. A tuned AI model with a published license, on macOS 14 and later.