The Speech-to-Text Industry in 2026: Market Size, Key Players, and Where It's Headed

For most of its history, speech recognition was a niche: dictation software for doctors and lawyers, call-center transcription, an assistant that sometimes understood you. That era is over. Speech-to-text is now a general-purpose input technology worth tens of billions of dollars a year, growing at close to 20% annually, and—in the most significant shift since recognition moved to the cloud—it is moving back onto the device.

How big the market actually is

Analyst estimates differ depending on what gets counted—pure transcription APIs, dictation software, voice assistants, or the whole “voice AI” category—but they agree on scale and direction. Fortune Business Insights puts the global speech and voice recognition market at $19.09 billion in 2025, rising to $23.70 billion in 2026 and a projected $104 billion by 2034. MarketsandMarkets, using a narrower definition, estimates $9.66 billion in 2025 growing to $23.11 billion by 2030. Grand View Research projects roughly $54 billion by 2030. Whichever definition you prefer, the compound growth rates cluster between 14% and 22% a year.

Individual company numbers make the scale concrete:

Microsoft paid $19.7 billion for Nuance—the company behind Dragon dictation and much of medical transcription—in a deal completed in 2022. It remains one of Microsoft’s largest acquisitions ever, made substantially for speech technology.
Otter.ai passed $100 million in annual recurring revenue in March 2025, built almost entirely on meeting transcription.
Deepgram raised a $130 million Series C at a $1.3 billion valuation in January 2026 to sell speech recognition as an API.
Wispr, maker of the dictation app Flow, raised $30 million in June 2025 and another $25 million that November; by May 2026, Bloomberg reported it was in funding talks at a roughly $2 billion valuation. That is venture pricing for an app whose entire product is typing with your voice.

The through-line: transcription itself is getting cheaper—in some cases free—while the money migrates to whatever sits on top of it: workflow, integration, and trust.

One terminology note

Speech-to-text (also called automatic speech recognition, or ASR) turns audio into written words. Text-to-speech generates spoken audio from written words—that is ElevenLabs’ core business, for example. The two increasingly share “voice AI” branding and sometimes the same underlying research, but they solve opposite problems, and their market figures are usually reported separately.

How the industry got here

Three eras explain the current landscape.

The desktop era (1990s–2010). Dragon NaturallySpeaking and its competitors proved dictation could work, but only after training the software on your voice, in specific applications, at prices that limited it to professionals with a strong need—medicine, law, accessibility.

The cloud era (2010–2022). Deep learning made recognition dramatically better, but the models were too large to run locally, so speech became a cloud service. This is the era of Siri, Google voice search, Alexa, and per-minute transcription APIs. Accuracy improved for everyone; in exchange, audio routinely left the device, and the economics favored a handful of companies that could afford to train and serve the models.

The open-model era (2022–present). In September 2022, OpenAI released Whisper—a recognition model trained on 680,000 hours of multilingual audio—as open source. Suddenly, near-state-of-the-art speech recognition was a downloadable file rather than a service. Competing open and open-weight models followed from NVIDIA, Alibaba’s Qwen team, and others, while runtime projects optimized them for consumer chips, including the Neural Engine in Apple silicon. The cost of good transcription collapsed toward the price of the hardware you already own.

The market is a stack, not a category

“Speech-to-text company” now describes at least five different businesses, and most players occupy only one or two layers.

Swipe horizontally to see the full table.

The layers of the speech-to-text industry
Layer	Who competes there	How they make money
Models	OpenAI (Whisper), NVIDIA (Parakeet, Nemotron), Alibaba (Qwen3 ASR), Google, Apple	Mostly indirectly—open models drive platform, hardware, and API adoption
Cloud APIs	Deepgram, AssemblyAI, OpenAI, Google, Microsoft/Nuance, Amazon	Per-minute or per-token inference pricing; enterprise contracts
Device runtimes	Apple (Core ML, Speech framework), Argmax (WhisperKit), FluidAudio, whisper.cpp	Hardware sales, developer licensing, or open source
Operating systems	Apple Dictation and Voice Control, Windows Voice Access, Android/Pixel voice typing	Free with the platform; competes on integration and defaults
Applications	Otter, Wispr Flow, Rev, meeting tools, medical scribes, Mac dictation utilities	Subscriptions, enterprise seats, or free products that support a broader business

Value is migrating upward. As open models push raw transcription toward commodity pricing, the defensible ground is the application layer: who owns the moment between deciding to speak and seeing correct text land in the right place. That is why a dictation app can raise money at a $2 billion valuation while the model it runs on is free.

The big shift: from cloud default to on-device option

The most consequential change in the industry right now is architectural. For a decade, “good speech recognition” implied a network connection. That assumption is breaking, for three reasons that arrived together:

Models got efficient enough. Distilled and quantized versions of Whisper-class models transcribe faster than real time on a consumer laptop.
Consumer chips got neural hardware. Every Apple silicon Mac ships with a Neural Engine; recognition that once needed a server rack now runs on the machine on your desk.
Privacy became a purchasing criterion. After years of stories about assistant recordings being reviewed by contractors, “your audio never leaves the device” became a feature people look for—especially journalists, lawyers, and doctors, whose recordings are precisely the ones that should not sit on a third-party server.

Platform vendors have moved with it. Apple’s built-in Dictation processes speech on-device in supported configurations and works anywhere text can be typed; Notes and Voice Memos transcribe recordings locally on Apple silicon; and Apple’s 2026 foundation-model research reports evaluators preferring its new dictation model over the previous system by 44.7% to 17.6% in Apple’s own side-by-side testing.

Below the platform layer, a cottage industry of independent utilities packages open models into system-wide dictation—a global hotkey, a small overlay, text pasted into whatever app is active, no audio uploaded. On the Mac these range from open-source projects to paid apps such as MacWhisper and Superwhisper to free ones such as our own Turbo Whisper, which wraps four open models (Whisper Large V3 Turbo, NVIDIA’s Parakeet and Nemotron, and Qwen3 ASR) behind a single Option–Space shortcut. The notable thing about this layer is not any one app; it is that a category which cost hundreds of dollars per seat in the Dragon era, and per-minute fees in the cloud era, is now often free, because the expensive part—the model—was given away.

Cloud recognition is not going away: servers still run the largest models, update centrally, and handle multi-speaker meeting audio and organization-wide search better than a laptop can. The change is that the cloud is now a choice with trade-offs—latency, retention policies, per-minute cost—rather than the only place good recognition lives.

What speech recognition still gets wrong

The industry’s marketing has outrun its measurement, and a few caveats belong in any honest account.

“Three times faster than typing” comes with fine print. The figure traces to a 2016 Stanford-led study that measured 161.2 words per minute by voice against 53.46 on a phone keyboard—prompted phrases, quiet room, mobile keyboard. Against a competent typist on a full keyboard, in a noisy room, with correction time included, the advantage shrinks and sometimes disappears. Speech wins most clearly for drafting at length, for sore wrists, and for hands-busy situations.

Accuracy is uneven across speakers. A 2020 study in PNAS found the major commercial recognizers of the time made nearly twice as many errors for Black speakers as for white speakers—a word error rate of 0.35 versus 0.19. Models have improved since, but the underlying cause—training data that underrepresents many dialects, accents, and acoustic conditions—has not been eliminated, which is why projects like Mozilla Common Voice exist. A headline benchmark number says little about how a system performs for any particular voice.

The errors that matter are not evenly weighted. Modern systems rarely mangle ordinary prose; they stumble on proper nouns, numbers, technical vocabulary, and overlapping speech. For casual notes that is cosmetic. In a quote, a contract, or a medical record, one wrong surname or a dropped “not” outweighs a page of perfect commas—which is why regulated fields still pair recognition with human review.

“On-device” requires reading the fine print too. An app may recognize speech locally but send the transcript to a cloud model for cleanup; a local transcript may still sync through a backup service; different features of one product can have different privacy behavior. The claim is checkable—a network monitor and the vendor’s documentation will tell you—but it has to be checked per feature, not per brand.

Where it goes from here

Three open questions will shape the next few years of the industry:

Does voice become an input layer rather than a product? The fastest-growing dictation tools don’t own a destination app—they type into everything: email, browsers, code editors, AI chat prompts. If that pattern holds, speech competes with the keyboard rather than with other transcription apps, and the addressable market is everyone who types.
Do users keep choosing models? Today’s independent utilities expose model menus the way early local-LLM tools did. History suggests mainstream products eventually hide that complexity behind automatic routing—the right model for the device, language, and task.
Can platforms absorb the category? Apple, Microsoft, and Google ship increasingly capable free dictation with their operating systems. Independent developers survive where they can move faster on models, interaction design, and niche workflows—the same squeeze that played out in screenshots, clipboard managers, and window management, categories where good independent utilities have nonetheless persisted for decades.

The bottom line: speech-to-text has crossed from feature to infrastructure. The models are increasingly open, the hardware to run them ships in ordinary laptops, and the money has moved up the stack to workflow and trust. The interesting competition is no longer over whether a machine can turn speech into text—it can—but over the few seconds between deciding to speak and seeing the right words land in the right place, without your voice traveling any further than it has to.

The speech-to-text industry in 2026: how big it is, who’s in it, and where it’s going