The dictation and speech-to-text market, product by product.

Dictation is a crowded market with a clear weak spot. There are far more products than most buyers expect, they cluster on Mac and phones, and the problem people raise most is recognition accuracy, not price or privacy. Accuracy in quiet rooms is largely solved: the engines we tested are within a few points of each other. It breaks down when other people are talking nearby, and it breaks down differently for each engine. For most people the free tool built into the operating system is the right starting point, and the data below shows where it stops being enough.

The market is dozens of apps and models, and Mac is better served than its share suggests

The map holds eleven segments. The segments, with product counts, are dictation apps (79), open models and engines (64), speech-to-text APIs (50), transcription apps (37), meeting note takers (36), medical and professional dictation (33), voice assistants (19), voice keyboards (17), built-in dictation (13), voice notes apps (13) and browser extensions (9).

Dictation apps skew to desktop: 66 of 79 run on Mac, 43 on Windows and 24 on Linux, while only 20 run on iPhone and 13 on Android. Voice keyboards are the mirror image, with 14 on Android and 8 on iPhone, but only 2 on Mac. Transcription is a web business (26 of 37 products). Medical and professional dictation has no Mac products in the map at all, and 16 of its 33 products run on iPhone.

The desktop numbers show a mismatch. Windows ran 76.41 percent of desktop page views worldwide ¹, macOS 16.74 percent ² and Linux 4.54 percent ³. Yet dictation apps list more Mac products than Windows products, so Mac users are better served than the page-view share implies. See the desktop OS share data.

Attention is going to chatbots and open-source engines, not to dictation as a topic

Researchers write about speech recognition at scale: 28,220 works in a year mention it ⁴, well above the level of a few years ago. "Speech-to-text" appears in 9,278 works ⁵, "voice assistant" in 6,728 ⁶, and "whisper speech" in 293 ⁷. The older word "dictation" grew least, at 3,139 works ⁸. Narrow topics stay small: "medical speech recognition" only 20 ⁹ and "speech recognition accessibility" just 4 ¹⁰. Details are in the research output data.

Public attention follows the chatbots. The English Wikipedia article on Google Gemini drew 1,093,862 page views in the past year ¹¹, Grok 1,037,826 ¹² and Microsoft Copilot 644,119 ¹³. Siri drew 307,597 ¹⁴, and carpal tunnel syndrome drew 191,983 ¹⁵. See the Wikipedia interest data.

Developers use open engines heavily. In the latest month of data, openai/whisper-large-v3-turbo was downloaded 6,094,388 times from Hugging Face ¹⁶, and the Apple-focused argmaxinc/whisperkit-coreml 11,924,373 times ¹⁷. On GitHub, whisper.cpp has 54,482 stars ¹⁸, and the open-source dictation app Handy 34,130 ¹⁹, against 167,850 for the general-purpose transformers library ²⁰. Data: Hugging Face downloads and GitHub speech repositories.

Downloads include automated pulls and stars are votes of interest, so neither counts users.

Users complain about accuracy far more than about anything else

We coded Hacker News comments about dictation for the problems they report ( dictation pain points). Recognition accuracy leads at 12.1 percent of comments ²¹. Everything else trails far behind, and the rarest problems, such as offline use, battery and noise, are barely mentioned:

Most comments report no problem, so read these as a ranking, not as the share of unhappy users. The data cannot show what people praise: we coded problems, not satisfaction, so we have no measured praise to report. Hacker News also skews toward developers, so it likely understates the complaints of people who dictate memos or medical notes.

Engines are close in quiet rooms and diverge when someone else is talking

We tested Whisper large-v3-turbo, Parakeet TDT 0.6B v3 and Apple SpeechTranscriber on a Mac with an M2 Pro, using English FLEURS recordings with background talk mixed in at set levels. Lower word error rate is better. Details are in the noise robustness data.

All values are percent word error rate. The findings are clear. First, with no noise or a quiet office all three engines are usable, and Whisper is lowest. Second, the order flips as the room gets louder: Parakeet barely moves from a quiet office to a busy one, while Apple's error rate rises sharply, and in a loud cafe Parakeet is clearly ahead. Third, when background talk is as loud as the speaker, every engine fails, and Parakeet's error rate is still lowest.

Published benchmarks agree on the rough picture but are messier. Across the on-device benchmark data, the median FLEURS error rate is 6.66 percent for Apple SpeechTranscriber ⁴⁷, 10.11 percent for Parakeet TDT v3 ⁴⁸ and 11.34 percent for Whisper large-v3-turbo ⁴⁹. The medians mix languages, and Apple's set is small. Counting punctuation and capitals, as real dictation must, the formatted accuracy data gives 10.62 percent for Apple ⁵⁰, 12.75 percent for Parakeet ⁵¹ and 12.94 percent for Whisper ⁵². Formatting costs every engine accuracy. The datasets hold 56 data points ⁵³ and 56 ⁵⁴, and the ranges behind these medians are wide.

Our test has limits. It uses English read speech, one Mac, background talk as the only noise, and the bare engines rather than the finished apps built on them. It says nothing about accents, other languages, headset microphones, phones or the cleanup that apps add afterward.

Free is the median price for dictation apps, and paid tools cluster in a narrow band

The dictation app price data shows a market with a free tier at its base. The median for the free plans is 0 US dollars ⁵⁵. A one-time purchase has a median of 44 US dollars ⁵⁶. Subscriptions run 15.99 US dollars per month billed monthly ⁵⁷ and 11.99 US dollars per month billed yearly ⁵⁸, so paying yearly cuts the monthly cost. The dataset has 44 data points from 19 sources ^{59 60}.

For developers, the speech-to-text price data shows engines priced in cents. Recorded audio has a median of 0.21 US dollars per hour of audio ⁶¹, and real-time streaming costs more, at 0.46 US dollars per hour of audio ⁶². Human transcription is a different product at 83.7 US dollars per hour of audio ⁶³, though that median rests on very few figures. The set holds 34 data points from 15 sources ^{64 65}.

These are list prices, not what customers pay, and they hide usage caps and feature differences.

Privacy rests mostly on what makers say, and the legal record is about voice assistants

The map records whether a maker says its product runs on the device. Among dictation apps, 39 of 79 do, and 39 are open source. Among voice keyboards, 9 of 17 say they run on device. Built-in dictation shows 1 of 13, because most platform tools use cloud engines for some languages or modes. Medical and professional dictation shows 0 of 33, and speech-to-text APIs 3 of 50, as expected for hosted services. Open models and engines are 63 open source out of 64 products. These are makers' statements, not audits, and "on device" does not guarantee that no audio leaves the machine for other features.

Privacy is a concern in user comments, at 2.4 percent of coded comments ²³, a leading concern among the problems. The enforcement record comes from assistants, not dictation apps. Apple agreed to settle a proposed class action claiming Siri violated users' privacy ⁶⁶. Google settled a class action over Google Assistant recordings and denied wrongdoing ⁶⁷. Amazon agreed to a civil penalty over alleged violations of a child privacy law relating to Alexa ⁶⁸. See the privacy settlements data and the voice assistants segment. These cases do not show how dictation apps handle audio.

What the data cannot show

The map counts products, not users, revenue or market share, so it cannot say which segment is biggest by money or people. Counts of platforms and on-device claims come from the products' own descriptions. Our accuracy tests cover English only and three engines. The complaint data comes from one forum. Prices are published figures, not transaction data. To choose a tool, start from the segment that matches your task, test a few products on your own voice and room, and weigh privacy claims against what the maker documents.