Machines now transcribe like humans, so value moves past the transcript

Microsoft's 5.1% error rate matches professional transcribers. With speech sold by the minute, companies must win on what follows the words.
Microsoft's conversational speech system now gets 5.1% of words wrong on Switchboard, a set of about 2,400 recorded telephone calls used to test speech software since the early 1990s. That is the error rate measured for professional human transcribers. The result, published on 20 August, cuts the company's error rate by about 12% in a year: its 2016 system stopped at 5.9%.
Parity matters less as a research trophy than as a price signal. When the best systems hear as well as people, transcription stops being a product a start-up can own and becomes a utility that the largest platforms meter by the minute. The money in speech moves to what happens after the words are on the page: understanding them, acting on them and fitting them to fields the general models handle badly, which is where our thesis on where AI value settles has pointed all along.
The race was short. Google put its own error rate at 23% in 2013, 8% in 2015 and 4.9% this May, measured on its own data. IBM reported 5.5% on Switchboard in March and, working with professional transcribers, set the human benchmark at the 5.1% that Microsoft has now matched. Three companies with cloud businesses reached the same level within six months, so none of them can charge a premium for it.
The price list already shows the result. Google sells its speech service at $0.006 per 15 seconds after a free first hour, or $1.44 for an hour of audio, and on 14 August added 30 languages for a total of 119. A start-up whose pitch is a better general transcription engine competes with that number, set by companies that treat speech as a way to sell cloud capacity and run it on chips they design themselves. Founders who build on these services keep their capital for what customers pay for: medical or legal vocabulary, structured records, the workflow around them.
Microsoft's own list of unsolved problems marks the openings: noisy rooms, distant microphones, accents and languages with little training data, and the step from recognizing speech to understanding it, which its researchers call the next frontier. On the harder CallHome set of telephone calls, IBM's best system still errs on 10.3% of words against 6.8% for people. Fanaura looks for speech companies whose edge sits in those gaps or in proprietary data from one industry, the part of the stack a family office investing in AI companies can hold for years without a platform pricing it away.
A fair counterpoint holds that parity on an old telephone benchmark says little about a crowded conference room, so the commodity argument runs ahead of the technology. That is right about the hard audio and wrong about the easy audio, which is already cheap and getting cheaper. The CallHome gap settles it. When machines close the distance to human transcribers on that set, the last technical reason to pay extra for generic transcription disappears.
The transcript has become the cheapest part of a voice product. Companies that sell it alone are selling what the platforms meter by the minute.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building speech products on proprietary data from one industry? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.