IA
FluidVoice: test the time you save after corrections
November 17, 2026
Why it matters
Dictation is appealing when your work involves explaining things: drafting an email, describing a bug, or giving a coding agent enough context to handle a task. Speaking gets the first draft onto the screen quickly. Whether it saves time depends on what happens afterward. Fixing a customer name is annoying; overlooking a changed instruction can affect the work itself.
Andrew Warner and Adam Brakhane make a convincing case for trying FluidVoice in this video. Their strongest demonstration is the live text preview, which helps a speaker see whether the microphone is capturing anything useful. Their experience also suggests that a free application can be comfortable enough to replace a subscription. It does not establish the same result for every machine, language, or workflow.
My evaluation would center on finished text: how long it takes to produce, how much editing it needs, and whether it preserves the intended meaning. The protocol below adds a repeatable comparison to the presenters’ demonstration. It is a proposed evaluation, with their hands-on experience attributed to them throughout.
Key points
- 00:34 Andrew raises cloud processing, waiting time, and interface design as reasons to switch. Each deserves a separate check, because removing a subscription does not automatically address the other concerns.
- 01:39 Model selection and keyboard shortcuts shape the setup. A comfortable shortcut and an appropriate language model are likely to matter more often than an impressive feature list.
- 02:12 The Wispr Flow demonstration highlights the delay after speaking stops. It is a useful illustration, although different sentences and a single attempt cannot establish a reliable speed ranking.
- 02:44 FluidVoice displays words during capture and identifies the destination application. That feedback can help catch microphone or focus problems before a long explanation is lost.
- 03:17 Adam emphasizes access to the source code. Inspection and modification are valuable options, with an associated maintenance burden for anyone who chooses to exercise them.
- 03:51 Custom prompts introduce an editing stage between recognition and insertion. That stage needs its own accuracy check, particularly when the dictated text contains instructions or commitments.
- 04:23 The custom dictionary targets recurring names and company terminology. Its value should be measured against difficult examples rather than assumed from the presenters’ confidence.
- 04:58 Adam chooses the faster model after trying alternatives. That is evidence about his preferences and workload, leaving other languages and hardware configurations open to testing.
- 05:31 Both presenters encountered installation progress that appeared stalled before completing. Treat this as an observation from their setup rather than a universal installation requirement.
- 06:38 Uncertainty about the active microphone brings the argument back to visible feedback. A dependable tool should make capture failures easy to notice.
How to apply
Give FluidVoice a representative workload
Check compatibility before downloading models. The official website currently specifies macOS 15 or later for the Mac application, with microphone and accessibility permissions required for capture and text insertion. Use the official FluidVoice website linked from the project.
Build a small evaluation set from the writing you actually do. Ten passages of two or three sentences are enough to expose recurring problems. Include a customer message, an explanation of a technical issue, and a task with explicit constraints. Use fictional details while preserving the complexity of your normal vocabulary.
Try a passage such as: “The budget is twelve hundred and fifty dollars. Do not deploy before Friday. Preserve the customer_id field in the integration.” Check the amount, the negation, the deadline, and the identifier independently. You may discover that speaking prose works well while typing identifiers remains faster.
If your work includes Brazilian Portuguese, test it directly rather than extrapolating from the English demonstration. The project lists Portuguese for Parakeet TDT v3 and Cohere Transcribe, while Parakeet TDT v2 is English-only. See the official model table.
Measure finished text instead of insertion speed
Compare FluidVoice with your current dictation tool and manual typing. Keep the microphone, destination application, and environment consistent. Repeat each task three times and alternate the order of the tools, so familiarity with a sentence does not consistently benefit the same option.
Record input time, waiting time after speech ends, correction time, and any meaningful errors left in the result. Distinguish cosmetic edits from changes to names, amounts, negations, and requirements. A missing comma and a reversed instruction should not receive equal weight.
Use a straightforward calculation: total time = input + waiting + review. For example, 35 seconds of speech, three seconds of waiting, and 40 seconds of editing produce a total of 78 seconds. These are illustrative numbers, not measured results. Compare the total with the time required to type an equivalent passage.
Also record failed insertions. Test an editor, an email composer, and whichever field you use to instruct agents. A tool that recognizes speech well but occasionally targets the wrong window adds a different kind of correction cost.
Check both stages of local processing
Speech recognition and AI cleanup are separate steps. FluidVoice supports local processing alongside optional cloud providers, so inspect the configuration of both stages. Once the necessary downloads finish, disconnect the network and repeat a complete task. This establishes whether the selected workflow works offline; it does not prove every aspect of its privacy behavior. The official FAQ explains the local and optional cloud distinction.
Source availability deserves a separate check if customization is part of the attraction. The application is GPLv3, while the repository describes Fluid Intelligence as a privately maintained local runtime. That distinction matters when assessing what you could inspect or maintain yourself. Read the project’s explanation.
Introduce cleanup only after testing recognition
Start with automatic rewriting disabled. Add recurring names to the dictionary and run the same passages again. Check whether spelling improves and whether similar-sounding ordinary words are mistakenly replaced by a saved name.
Then try a restrained cleanup instruction: “Fix punctuation and remove hesitation. Preserve names, numbers, negations, and technical terms. Do not add information or translate.” Compare the edited output with the intended message, paying particular attention to conditions and uncertainty.
For coding instructions, inspect verbs and permissions before submission. An instruction to investigate should retain that scope. For customer messages, check recipients, amounts, and promises. These checks belong in the measured review time.
Make the switch after a week of ordinary use
Use your normal microphone and working environment for a week. Track a handful of representative tasks each day, including sessions where speaking feels inconvenient or the surrounding noise makes capture unreliable.
I would switch if the total completion time consistently improved, consequential errors remained easy to spot, and the chosen configuration met my preference for local processing. A lower subscription bill is useful evidence in the decision, alongside the editing effort that remains every day.
Resources mentioned
- FluidVoice · Aplicativo gratuito; o endereço apresentado no vídeo redireciona para este site oficial.
- Código-fonte do FluidVoice · Repositório oficial, documentação e downloads; distingue o app aberto do runtime Fluid Intelligence.
- Wispr Flow (paid) · Comparado na demonstração; oferece modalidade gratuita limitada e planos pagos.
- Superwhisper (paid) · Alternativa usada por Adam; oferece modalidade gratuita e recursos Pro pagos.