Speech-to-Text Is Now a Commodity. Intent Is the New Moat.

Reson8 published a number this month that deserves more attention than it received. Its Rumble-2-Flash model transcribes an hour of audio for about 0.014 euros, which is roughly a cent and a half. Put that next to the price of almost any other computing service and the scale becomes obvious. An hour of speech contains thousands of words, dozens of speakers and a complete record of a meeting, and the processing cost is smaller than the electricity used to display the file. When a capability gets that cheap, it stops being a feature and starts being infrastructure, like compression or hashing.

The collapse did not happen in one step. It came from three forces arriving at once. Model architectures got more efficient at the same accuracy, so the same quality required less compute. Competition pushed providers to undercut each other on price per minute, which turned transcription into a race toward the floor. And demand grew enough to justify batch processing and specialized hardware, which lowered the unit cost further. None of these forces is exotic. Together they produced a market where the headline number is no longer accuracy but cents per hour, and where every vendor has to answer to that number.

For buyers, the practical consequence is that transcription is now a line item, not a strategy. Teams can switch providers with modest effort because the outputs are broadly comparable and the integration surface is small. That is good news for cost control and bad news for anyone selling transcription as a differentiator. When a component is interchangeable, procurement treats it as interchangeable, and the conversation moves to volume discounts and reliability rather than capability. The technology is better than ever, and it is also less interesting than ever, which is exactly what happens to every maturing commodity. That is not a complaint; it is a sign of success.

For builders, the implication is harsher. A product whose only value is converting audio to words is competing on a price that is already near zero. Margins compress, differentiation disappears, and the roadmap becomes a series of small accuracy improvements that users cannot perceive. The companies that survive that transition are the ones that move up the stack before the floor arrives. They stop selling recognition and start selling what recognition makes possible. That shift is not a marketing exercise. It is a change in what the product produces, and it decides which vendors still have a business in two years.

The other half of the story is size. Underdog demonstrated a speech model of roughly 32 megabytes, small enough to run on an edge device rather than a data center. Thirty-two megabytes is smaller than many photos and smaller than the installers people download without thinking. A model that size changes where inference can happen. It can live on a laptop, a phone or an embedded device, and it can process audio without a network round trip. The accuracy is not identical to a large cloud model, but for many everyday tasks it is close enough, and the deployment economics are completely different.

Edge inference matters for three reasons that compound. Latency drops because audio no longer travels to a server and back, which makes the experience feel immediate. Cost per request collapses because there is no per-minute API bill, only the device the user already owns. Privacy improves because the audio can stay on the machine, which removes an entire category of compliance question. Those three benefits reinforce each other. A local model that is fast, free at the margin and private is a different product from a cloud service, even when both produce the same transcript. The difference is invisible in a demo and decisive in daily use.

The architectural shift also changes who can ship voice features. A startup without a data center budget can embed a 32MB model and offer dictation in an application that would never justify a server fleet. An enterprise can deploy on locked-down machines where outbound audio is forbidden. A field worker can dictate without connectivity. The long tail of voice use cases, the ones too small or too sensitive for a cloud API, becomes addressable. Commoditization at the top of the market and miniaturization at the edge are two expressions of the same trend: speech processing is becoming ambient and abundant.

What the 32MB model does not change is the output. A small model still returns words. It does not decide which words were an instruction, which were context and which were a correction the speaker immediately abandoned. It does not add a greeting because the text is going into an email, or compress a ramble into three bullet points. Those tasks require a different layer, and that layer has nothing to do with model size. This is the boundary that matters for product strategy. Recognition is being solved twice, once in the cloud and once at the edge. Expression is still mostly unsolved, and it is where the remaining value sits.

Word error rate is the metric the industry reports and the metric users never feel. A transcript can score well on a benchmark and still take ten minutes to turn into a message. What users experience is post-edit effort: how much they have to change before the text is safe to send. That number is rarely published because it is harder to measure and less flattering. It is also the only number that predicts whether a voice feature gets used twice. Accuracy is a necessary condition for adoption. It is not a sufficient one, and treating it as the goal produces tools that demo beautifully and stall in daily work.

The reason is structural. Speech is not writing. A person talking produces filler, false starts, repeated clauses and sentences that change direction halfway through. They use pronouns for things that are visible in the room but absent from the text. They speak in one long stream where a writer would insert a paragraph break. None of that is an error in the recognizer. It is the nature of spontaneous speech. A literal transcript is therefore an accurate record of something that was never meant to be read. Handing it to a user and calling the job done is like handing someone a recording and calling it a document.

Consider a realistic example. A three-minute spoken draft of an email becomes a wall of text with four abandoned openings, two restarts and no paragraph structure. Cleaning it takes longer than typing the email would have. The voice feature saved two minutes of speaking and cost six minutes of editing, so the user stops using it and tells colleagues it does not work. The failure is not accuracy. The transcript may be nearly perfect. The failure is that the tool delivered the wrong artifact, a record instead of a draft. Every product that sells transcription as an end state runs into this wall eventually.

The fix is to change the target. Instead of asking whether the words are correct, ask whether the output is sendable. Sendable text has punctuation, paragraphs, a clear opening and a clear ask. It has the speaker's corrections applied rather than preserved. It matches the conventions of its destination, whether that is a chat message, an email or a prompt. That standard is harder to hit than a benchmark, and it is the only one that matters to the person holding the cursor. Products that adopt it stop competing on recognition and start competing on what they produce, which is a much more defensible position.

Intent understanding means recovering what a speaker meant, not just what they said. In practice, that requires several judgments at once. Which clause is the instruction and which is background. Whether a repeated sentence is emphasis or a correction. Whether the destination is a message to a colleague or a prompt to a model, because the two need different shapes. Whether a name is a person, a product or a typo. A recognizer is not asked to make any of these judgments, and a general-purpose language model is too slow and too context-blind to make them reliably at the cursor.

That difficulty is what makes intent a moat rather than a feature. Recognition improves through scale and data that everyone can buy. Intent improves through integration with the places where text is actually produced, the conventions of specific workflows, and the accumulated preferences of a user or a team. Those assets are not available in a public benchmark. They are built by being present in the editor, the inbox and the chat client, learning what someone means by make this shorter in each context. A competitor can license the same speech model. It cannot license the accumulated understanding of how a particular team writes.

It helps to think in layers. The bottom layer is recognition, which is now a commodity with a visible price per hour. The middle layer is structuring, which turns a stream of words into paragraphs, lists and correctly punctuated prose. The top layer is expression, which adapts tone, format and vocabulary to the destination and the person. Each layer adds value and each layer is harder to copy. Recognition can be swapped in an afternoon. Structuring requires product judgment. Expression requires presence across applications, a privacy model users trust, and enough usage to learn preferences. The moat lives in the top two layers.

Competitive dynamics follow from that stack. Model providers are strong at the bottom and increasingly at the middle, but they are structurally distant from the top because the top is cross-application and local. An assistant inside one product can learn that product well and nothing else. An input layer sits underneath every product, which gives it a different view and a different set of habits to learn. That position is unglamorous and hard to dislodge. It is the kind of advantage that does not show up in a launch demo and does show up in retention, because the tool becomes part of the muscle memory of writing.

Expression intelligence is the practical name for that top layer. Its output is not a transcript but a piece of text a person can send without rewriting. Concretely, that means filler removed, restarts resolved, punctuation and paragraphing added, and the result shaped for its destination. The user still reviews and can still edit, but the edit is short. The measure of success is not words per minute. It is the ratio of inserted text to surviving text, and how often the first draft is good enough to send after one read. That is a testable product claim rather than a vague promise about understanding.

The layer has to work everywhere, because work does not stay in one window. System-level voice input on Mac and Windows puts the same flow behind one hotkey in Gmail, Outlook, Slack, documents, Cursor and any prompt box. Support for more than one hundred languages matters for teams that think in one language and write in another, and it prevents the tool from becoming useful to only part of a company. Those two properties, universal coverage and language breadth, are what turn expression from a demo into a daily habit. A feature that works in one app is a novelty. A feature that works at every cursor is infrastructure.

The workflow that makes it usable is deliberately small: speak, polish, type, edit. The user holds a hotkey, speaks naturally, releases, and the polished text appears at the insertion point. Corrections are treated as edits rather than preserved as artifacts. Privacy is designed into the same path, with processing local by default, zero retention and independent attestation such as SOC 2 Type II, because dictation contains the material people are most careful about. For organizations, a Team plan distributes shared vocabulary and formatting conventions without pooling individual drafts, which lets a group converge on consistent output without surveillance. That balance is what makes a team rollout sustainable.

For product leaders, the evaluation is straightforward and unforgiving. Measure how much of the inserted text survives, how many tasks finish faster, and how many drafts stop slipping to the next day. Those numbers are the real return on a voice investment, and they are not visible in a recognition benchmark. The strategic conclusion follows from the economics. When transcription costs a cent an hour and fits in thirty-two megabytes, no one builds a durable business on hearing alone. The durable position is producing text that people trust enough to send, and that position belongs to whoever owns the expression layer.