Voice is the fastest way into AI: what we measured when we put voice in front of it

The argument, and what we can and cannot show

Typing into a blank box makes you tidy your thought before the AI sees it. Talking does not. You say the half-formed version, the aside, the thing you are not sure about. That is the case for voice, and it is our view, not a measured result. We have not timed typing against speaking, so this page makes no claim about how many times faster one is.

What we can show is what happens once voice is in the loop: how quickly an answer must start, what it costs by the minute, and what breaks. Those we measured.

An example of unstructured input

On 27 October Rishabh wrote one file of notes that jumped between subjects with no order. In one run of about 35 lines it covered the Android app, the site header, how Arjun and Meera sound, the logos, the podcast feed text, internships, the admin screen, and how to stop writing the same article many times. It said, among other things, that the Intelligence page had "changed to voice, then back to Simli" and asked whether the old working version had been checked.

The file's header says the planning job (FABLE-11) reads it, and the direction document of 28 October is the result: one written plan for every product and lane. That document also records a request for "content on how voice speeds ideas into AI". That is this page. The order was made after the thought, not before it. If you dictate your own notes, the same thing applies: do not edit while you speak; sort afterwards.

We cannot tell from the file whether those notes were spoken or typed. If you build this way, the useful part is the same: write or say the whole thing first.

Speed of the first sound

For a conversation, the number that matters is the time from the moment you stop speaking to the first sound of the answer. We measured it on 26 October from one Mac, using the median across turns.

Our own build is the 4.9-second row. The 9.5-second row is the one we parked.

Interruption also differs. If you speak over the answer, OpenAI stopped in 0.31 seconds, ElevenLabs in 0.51 seconds and Simli in 0.84 seconds. For Hindi, OpenAI handled both directions on every turn; ElevenLabs needed a second agent and merged two Hindi questions into one.

One more finding corrected our own memory. A note said the face service had worked better before. When we tested it, the first word came 4.8 to 10.8 seconds after the person stopped, and the speech-to-text cut sentences in half and waited. The old version also had our own server in the loop, so it was never faster than the current build.

What we picked, and what it costs us

We built the OpenAI Realtime mini route behind a switch. The next step is to turn it on for Rishabh's own account only and run a phone test. At full use, our estimate was ₹108 a month per paying user. We give up three things: Claude as the voice's brain, nine other languages Sarvam covers, as a promise, and the face. The knowledge of the product stays, because the industry prompt and the day's brief are put into the session. The 5-minute phone test on a real network is the part we have not proven.

Where voice went wrong in production

Voice failures are different from text failures. They are heard, not read.

• The accent changed in the middle of an episode. The English voice drifted to another accent. The text chunks sent to ElevenLabs were 2,500 characters with no language set, and the limit is 2,000. They are now 1,800 characters or fewer, the language is fixed to English, and each episode uses one seed. We found it on 23 October.
• A wrong version was voiced. On 22 October both 09:15 script attempts failed. The fallback, which reads the headlines, was voiced about ten seconds after the second retry was sent. It was published and then pulled, and the Hindi edition had already been translated from it.
• A number was misread. A check read "web3" as an unsourced number "3" and failed a good script. The same run had scripts of 5,879 and 6,170 characters against a 7,200 minimum.
• The credits do not add up. The ElevenLabs plan has 150,000 characters a month. A daily episode is about 7,000 characters, which is about 210,000 a month. So a guard drops the episode to Sarvam when the credits run low.
• A page must not make audio by itself. Read-aloud on a page is made only when someone presses play, and it is stored so the next listener gets it free. A news page's listen button must never create a new podcast.

What we do about it

None of these were solved by a better model. They were solved by limits and checks around the voice: a character cap on every chunk, a language fixed in the request, a check that every spoken number is in the source, a stored copy before the second listener, and a switch that can turn a paid voice route off.

If you add voice to a product, we would start with three things:

1. Measure the time to first sound before you build anything else. It ranged from 1.0 to 9.5 seconds across the five set-ups we tried.
2. Price a speaking minute before you give it away. The same minute cost us between ₹1.8 and ₹16 depending on the service.
3. Keep a person's thought whole. Let people say it first and sort it afterwards.

Questions people ask

Is voice really faster than typing?

We have not measured it, so we do not claim a figure. What we measured is how fast an AI voice answers: 1.0 to 1.3 seconds at best, 9.5 seconds at worst.

Which voice service answered fastest?

On 26 October the OpenAI Realtime full model answered in 1.0 seconds and the mini model in 1.2 seconds, against 1.3 seconds for ElevenLabs.

Which was cheapest?

OpenAI Realtime mini at about ₹1.8 for a speaking minute. Our own Sarvam and Claude build was about ₹3, and ElevenLabs about ₹16.

Why did you park the talking-face service?

It took 9.5 seconds to the first sound, and it cost about ₹8 to ₹12 a minute. We built a face that runs in the reader's browser and costs nothing per view.

Can I try it?

The Intelligence page and the face maker are public. The real-time route is behind a switch and not yet open to visitors.

Sources

• Intelligence
• Face maker
• The podcast
• Our own measurements and logs, read on 31 October 2026: the 26 October voice comparison, the 27 October notes, the direction notes of 28 October and the AI-LOG entries of 22 and 23 October.