How Ace Coach Evaluates Speaking: Methods & Limits

By the Ace Coach team, built with the best speech coaches · Updated October 25, 2026 · How we write

The short version

Code measures. Pace, filler words, pauses, pitch, how your sentences end and, with the camera on, where you look. These are calculated from your audio and video, the same way every time.
AI judges. An AI review reads your transcript together with those numbers and decides what they mean for this call: whether you answered the question, whether your point came through, and which habit is worth working on first.
Where they meet. The AI is given the measured values instead of guessing them. It does decide a few things that feed into numbers, like whether a “like” was filler, and it scores each area; code then combines those area scores into your overall score.

What happens after you hang up

Transcription. Each of your turns is transcribed word by word, with a start and end time for every word, by a speech recognition model (Whisper large-v3). It's told to keep “um” and “uh” instead of tidying them away.
A coverage check. If a turn comes back with too few words for how long you were talking, it's transcribed again. If it's still thin, the review says it was only partly transcribed, and pace and filler rates for the whole call are left out rather than guessed.
Speech and silence. A voice activity detector (Silero VAD) marks where you were speaking and where you paused.
Voice. Pitch and loudness are measured every 20 milliseconds with Praat, a standard tool in phonetics research.
Camera, if it was on. Face landmarks are tracked frame by frame (MediaPipe) to measure eye direction, blinking and head movement. With the camera off, the review covers your audio only.
The review. The AI review reads your transcript and the measured numbers and writes your feedback. It's usually ready a few minutes after the call, depending on how long you talked.

Example of a pace note (illustrative): “Pace: 182 words per minute, above the 125–170 range. You sped up most while explaining the budget, at 1:42.”

What code measures

Measure	How it's calculated	Range the review starts from
Pace	Words per minute over your speaking time, from the first to the last word of each turn. Fillers aren't counted as words.	About 125–170
Filler words	“Um”, “uh”, “er” and “ah” are counted by spelling. Words like “like”, “so” and “you know” are flagged, and count only if the review judges them filler in context.	2 or fewer a minute
Pauses inside a thought	Code finds every silence of half a second or more. The review decides which ones broke a thought; pauses between sentences and deliberate pauses before a key line don't count.	0–1 per answer
Pitch variety	How much your pitch moves, averaged over 10-second windows (pitch variation quotient).	0.20 or more
Statement endings	Finished statements whose last syllable rises 3 semitones or more above your usual pitch, so they sound like questions.	20% or fewer
Trailing off	Phrase endings where the last half second is 3 dB or more quieter than the rest of the phrase.	20% or fewer
Eye contact	Share of the time your eyes are open and not looking down or more than 15° to the side. Camera only.	70% or more
Blinking	Blinks per minute while you speak. Camera only.	30 or fewer
Talk share	Your speaking time compared with the other person's, on practice calls.	Interviews 50–75%, other calls 35–55%

These ranges are coaching guides, not research norms. They're where the review starts, not a verdict: it can decide a fast pace is fine because you were clear, or that a pace inside the range still dragged. Some measures are less settled than others: blink counts and trailing off are treated as hints the review checks, and the pitch variety range is provisional. Talk share is shown for context and is never set as a goal.

What the AI judges

Whether you answered the question, and how early you got to it.
Your opening and close: whether your point comes in the first sentence and whether you end on a clear point or next step.
Specifics: concrete facts, numbers, names and examples.
Structure: only the parts a good answer to this particular question needs, shown as “2 of 3 parts”, never a fixed five-part checklist.
Hedges and vague words: code flags candidates like “maybe” or “stuff”; the review decides which ones weakened what you said.

Suggested rewrites are built from what you said and what you told Ace about yourself. Where a detail is missing, they leave a blank for you to fill in instead of inventing one. Still read a rewrite before you use it. Quotes in your feedback are matched to the exact words in your transcript.

How goals work

Ace picks a few habits from your calls and tracks them call after call. A goal counts as reached only after several good calls, not one lucky one. Very short answers don't count toward goals, and calls with almost no speaking aren't scored.

Known limits

Lens or screen. The camera can't tell whether you're looking at the lens or at the screen, so both count as eye contact.
Transcription errors. Speech recognition makes mistakes, more often with strong accents, names, jargon or a noisy room. Word times can be off by a few tenths of a second.
Loudness reflects your setup. How loud you sound depends on your microphone and how far you sit from it as much as on your voice. If your browser is leveling the sound automatically, loudness isn't measured at all.
English ranges. The pace and filler ranges are for English.
Not clinical. Ace doesn't read emotions, detect lying or assess stuttering or other speech disorders. It isn't a substitute for a speech-language pathologist.
Camera measures need a clear view. If your face is out of frame, too dark or turned away, those stretches are left out, and some measures may be missing.
The AI can be wrong. Feedback points to the moment it's about, so you can listen back and decide for yourself.

Your recordings

The privacy policy explains how recordings are processed, which service providers handle them and how long they're kept. Recordings are not sold or used to train AI models. For how this site's articles are written and sourced, see how we write.

Found something that doesn't match what you saw in the app? Email [email protected].

How Ace Coach evaluates speaking.