Typing a 250-word prompt takes me three to five minutes. Speaking it takes about one. Most people speak at 120-160 words per minute and type at 40-70, the gap has always existed, but it used to be eaten by transcription errors and cleanup. Modern speech models closed that gap, and LLMs on the receiving end removed the last excuse. Switching to voice dictation has been the single biggest productivity improvement in my AI workflow this year.

Why voice and LLMs fit together #

Dictation historically failed because the output went straight to a human: every disfluency, every “um, actually scratch that,” landed on the page and had to be fixed by hand. When the receiving end is a model, that problem disappears. LLMs tolerate rambling, restore structure, and fix grammar as a side effect of doing their job.

This changes how you prompt, not just how fast. Instead of carefully crafting a prompt sentence by sentence, you think aloud and let the model organize your ideas. The interaction shifts from “compose, then submit” to “brain-dump, then refine”, and the brain-dump usually contains more context than the carefully typed version would have.

Better prompts, not just faster ones #

The less obvious benefit is prompt quality. Typing has a cost, so you compress: you drop background, skip constraints, and hope the model infers the rest. Speaking is nearly free, so you elaborate. My dictated prompts are consistently longer and carry more of the surrounding context than anything I would have typed, and that matters, because prompt research consistently shows that specific, context-rich prompts produce better outputs. Recent work on domain-specific tasks found that longer, more detailed prompts improve model performance where precise reasoning is required (Liu et al., 2025). To be fair, nobody has run a controlled study of spoken versus typed prompts yet. But the mechanism is simple: speaking lowers the cost of including detail, and detail is what LLMs reward.

Best practices #

• Speak thoughts, not syntax. Don’t dictate punctuation or formatting. Describe what you want in plain speech and let the LLM structure the output.
• Don’t self-correct mid-sentence. Stopping to fix a misheard word breaks your flow and rarely helps. Keep talking; fix it in review.
• Use it for prose-shaped work. Prompts, PR descriptions, commit messages, documentation, issue tickets, code review comments, architecture notes. Writing code by voice is a different discipline with much smaller returns.
• Match the model to the content. Cloud models for maximum speed and accuracy; local models when the content shouldn’t leave your machine. Good tools let you switch per context.
• Feed it your vocabulary. Project names, domain jargon, teammate names. Most tools accept a custom dictionary or context prompt, and it dramatically cuts correction time.
• Always review before sending. Total task time includes editing, the research on interactive dictation is clear that raw entry speed isn’t the whole story.

Tools #

My recommendation is OpenWhispr: open source (MIT), cross-platform, and flexible about where processing happens. You can use cloud models for fast, accurate transcription, or run Whisper and Parakeet locally with your audio never leaving the machine, with bring-your-own-key support for the cloud side.

Worth knowing about:

• Wispr Flow: the most polished commercial option, cloud-only.
• Superwhisper: custom modes per task (code, email, prompts) with local and cloud models; macOS, Windows, iOS.
• VoiceInk: fully local, privacy-first; macOS (Apple Silicon) and iOS.
• Handy or whisper.cpp: the open-source route if you want an offline stack you fully control.

And if you feel fancy: this is a genuinely easy tool to build yourself. A global hotkey, a local Whisper model, and a paste into the active window is the whole product. An afternoon with Claude gets you a custom dictation tool adapted exactly to your workflow (your vocabulary, your formatting rules, your cleanup prompt), and you avoid another subscription.

What the numbers say #

The strongest evidence is on raw text entry. A Stanford study (Ruan et al., 2016) found speech input 3.0× faster than keyboard typing for English, with a 20.4% lower error rate, and that was with 2016-era speech recognition.

More recent research fills in the workflow around it. Toward Interactive Dictation (ACL 2023) studied combining dictation with spoken editing commands and quantified the accuracy-versus-latency trade-off that dictation systems still navigate. StepWrite (CMU, UIST 2025) showed that LLM-guided voice composition significantly reduces cognitive load for long-form writing in a 25-participant study. And Moonshine (2024) demonstrated real-time local transcription at 5× less compute than Whisper tiny with no accuracy loss, the reason fully offline dictation is now practical on ordinary hardware. There’s even work prompting Whisper itself to do language understanding beyond transcription (Interspeech 2024), speech models are becoming full participants in AI pipelines, not just input devices.

End-to-end gains are smaller than the 3× entry speedup, because review and editing time doesn’t shrink. From my own experience, not a research finding, a realistic expectation for a writing-heavy AI workflow is 15-30% overall, with much larger gains on the days dominated by prompts, docs, and reviews.

References #

• Ruan et al., Speech Is 3x Faster than Typing for English and Mandarin Text Entry on Mobile Devices (Stanford, 2016)
• Li, Eisner, Pauls, Thomson, Toward Interactive Dictation (ACL 2023)
• El Alaoui et al., StepWrite: Adaptive Planning for Speech-Driven Text Generation (CMU, UIST 2025)
• Jeffries et al., Moonshine: Speech Recognition for Live Transcription and Voice Commands (Useful Sensors, 2024)
• Li et al., Prompting Whisper for QA-driven Zero-shot End-to-end Spoken Language Understanding (Interspeech 2024)
• Liu, Wang, Willard, Effects of Prompt Length on Domain-specific Tasks for Large Language Models (2025)