We’re engineers, not a research lab.
We didn’t arrive at Speaking DNA through a stack of papers. We built real speech pipelines, ran them on thousands of real sessions across seven languages, got plenty of things wrong, measured what actually happened, and fixed them. This page is an honest look at how the feature really works.
The measurement
Six dimensions, and exactly how we get each one
No black box. Some of these come from real acoustic measurement of your voice; others from analysing what you actually said. We tell you which is which.
Speaking pace and pause patterns — when audio is available, blended with the real speaking-vs-silence ratio measured from the waveform.
Phoneme-level scoring from a dedicated speech engine (Azure Speech). It runs on assessments and voice checks, where we actually have your audio — not on every casual chat.
A weighted mix of hesitation markers, pace variance and pausing — the things listeners actually perceive as "flow".
Grammar correctness per turn, cross-checked against a model-graded score, plus the five error patterns that show up most for you.
How varied and precise the words you reach for are — the unique-word rate across what you actually say.
A proxy — not a clinical measure — built from response latency, fillers and self-corrections. We label it an estimate, because that is what it is.
Acoustic signals (pitch, energy, pauses, voice quality) are real digital-signal-processing measurements from your audio. Speaking rate and the proficiency scores are derived from your transcript and a language model. We never blur the line between the two.
Your Voice Fingerprint
Six signals we read straight from your voice
When we have your audio — on speaking assessments and periodic voice checks — we run real signal processing over the waveform. These are the raw acoustic signals behind the six dimensions, captured and anchored against your very first session so we can see how you change.
Your fundamental frequency — average and how much it moves. Flat pitch reads as monotone; range reads as expressive. Measured with Praat.
Steadiness of your voice across a phrase, derived from period-level variation. Treated as an internal signal, not a clinical verdict.
Words per minute, from your transcript and the clip length. Too slow can mean searching for words; too fast can hurt clarity.
How much you project, and how much it varies. Measured as frame-by-frame loudness from the audio itself.
Hesitation markers and where you pause. Mid-sentence pauses signal formulation effort; pauses at boundaries are natural.
Amplitude consistency across the signal, from Praat. A vocal-characteristics signal we use internally — never sold as a "confidence score".
We capture these where the audio actually reaches us — assessments and voice checks — and hold them against your baseline. We don’t pretend to measure your pitch or energy from a casual live chat that never leaves your device.
Trial, error, fix
The things we got wrong first
The feature you use today is the version that survived contact with real learners. Here are four of the corrections that shaped it.
When the numbers were too kind, we tore them up
An early version of Speaking DNA produced strands sitting at 66–100% while the underlying assessment said 12–30%. The scores felt great and meant nothing. We rebuilt the pronunciation strand on a real phoneme-scoring engine so the number you see reflects how you actually sounded — even when that number is lower than you hoped.
A model that understood you but wrote down nonsense
Our voice model replied perfectly to a beginner saying "Ik kijk vaak" — yet the separate transcription layer wrote "Ikkaikvak". Garbage transcripts quietly broke the displayed text, the grammar corrections, and the DNA itself. We tested transcription models on real beginner, accented speech and switched to ones that hold up on short, imperfect utterances — with an environment switch to roll back in minutes if a model regresses.
Refusing to score what we did not hear
Live conversations stream peer-to-peer to the voice model — for privacy and latency, the raw audio never reaches our servers. That means we genuinely cannot measure your pitch or pauses from a casual chat. Rather than fake it, we "pin" the acoustic strands: they hold their last real value instead of drifting on guesses. Acoustic measurement happens on assessments and periodic voice checks, where we actually have the audio.
Tuned against real sessions, not a lab
Energy thresholds, pause-length cutoffs, minimum-duration gates, smoothing rates — none of these came from a paper. They came from running the pipeline on real sessions across seven languages and adjusting until the output matched what a human would say about that recording. Every dimension updates as a slow moving average, so one odd session can never whiplash your profile.
How we hold the line
The rules we build by
Where we have your audio, we run genuine signal processing — pitch, energy, pauses, voice quality — not a language model guessing how you probably sounded. Where we only have text, we say so.
Acoustic analysis happens entirely in memory. The only temporary file we ever write is the one a speech library strictly requires, and it is deleted the instant analysis finishes. We store numbers, never recordings.
We compute voice-quality signals like jitter and shimmer, but we treat them as internal heuristics — not as a "science of confidence". We will not dress up a vocal-health metric as a proficiency score.
Models, thresholds and scoring methods sit behind switches. If a new transcription model or scoring change makes things worse for real learners, we roll it back in minutes — no redeploy, no waiting.
Grounded in the field, honest about the limits
Our choices line up with what second-language research consistently finds: that pronunciation feedback works, that fluency is really about pace and where you pause, and that intelligibility matters far more than erasing an accent. We lean on that evidence — but the engineering, the testing, and the calibration are ours.
And where the science gets thin, we stop. We don’t claim to read your emotions from your voice or score your character. We measure speaking, carefully, and we tell you the truth about what the numbers mean.