About ImprovTalk

What it is

ImprovTalk is a voice-first conversation coach for iPhone. You hold a button and talk to an AI that answers back in real time, in a scene you pick — a café queue, a gym, a networking room — and afterwards you get a score with the working shown.

It exists because conversation is a motor skill and real life is a terrible place to practise one. You get a single attempt, no feedback on what went wrong, and a social cost for every fumble. Nobody would learn an instrument that way.

Who makes it

ImprovTalk is built and run independently, not by a large company. There is no content team and no outsourced blog — everything on this site is written by the people building the app, which is also why it can tell you exactly which line of the scoring pipeline a number comes from.

Support goes to a person, not a queue: davidgadyan92@gmail.com. Every email is read.

What the method is built on

The conversation technique the app coaches — and everything in the writing here — comes from a distilled corpus we assembled: fourteen curated sources on approach, framing, humour, observation and handling a no, classified into rules rather than kept as prose. The single largest source was a set of several hundred short coaching videos, topic-tagged so recurring principles could be separated from one-off opinions.

We do not name those sources on the site, and that is deliberate. The distillation is ours; citing the material it came from would imply endorsements nobody gave.

On top of that sits the content you actually practise against:

• 70 live practice scenarios across 13 venues, and 210 simulator scenes across 19.
• 195 nationalities, each with a phonetic greeting and a joke, which shape who the AI partner is and how she sounds — they change her accent, not the language she speaks.
• A reference library of 2,766 items across 25 catalogues, for when you run out of things to talk about.

How the score is actually calculated

Most apps in this category give you one number and no way to check it. Here is ours, in full.

The words are graded against a fixed rubric on three axes: Fluency (40%), Confidence (35%) and Improvisation (25%).

The delivery is measured from the recording itself — not inferred, measured — and combines depth (15%), intonation (15%), pace (25%), pauses (18%), fewer fillers (17%) and steady volume (10%). Pitch comes from acoustic analysis, pace from word timings, filler density from a fixed marker list counted per hundred words.

The two halves are blended, with delivery counting for 25% of the headline number. The method pages explain what to do with the result.

What we do not claim

This is the part most product pages leave out, and it is the part worth reading.

• The score is not fully deterministic. The measured delivery half is — the same recording produces the same numbers. The rubric half is a model judgement made the same way every time, which is not the same thing as identical every time.
• The app’s interface is English only. Conversations run in English, Spanish or Russian. The menus do not.
• We publish no ratings or testimonials. There are none yet. Inventing them is both a policy violation and a lie.
• No numbers we have not counted. Every figure on this site is counted from the app’s own data files. When a count changes, the site changes — we corrected one downward recently rather than leave it flattering.

How the AI partner behaves

She has agency. She can be guarded, uninterested, or done with the conversation, and a soft no or a hard no is a scored outcome rather than a failure state. Noticing one early scores better than talking through it.

The app will not coach coercion, and it is built to decline it. That is a product decision, not a legal disclaimer — reading a no is the actual skill.

Privacy, briefly

Your voice is used to score the session and nothing else. The microphone is not active outside a session screen, no contacts, calendar, photos or location are collected, and the app ships no advertising SDKs. The privacy policy has the detail, and support covers export and deletion.