Plenty of measurable skills have a number attached. Reading has grade levels, language ability has CEFR bands, typing has words per minute. Everyday spoken communication has no equivalent that most people ever encounter, and the usual feedback loop is still a coach in a room giving an opinion.
We built the Wellspoken Index to change that, and this week we published a dataset it produced: 16,928 candidate answers, drawn from a random sample of 1,000 AI-led mock interview sessions.
The median interview answer runs 17 words. The longest single answer in the sample ran 2,010 words, roughly thirteen minutes without stopping.
The reason this could not be done before
The obstacle was never the microphone. It was that judging speech meant asking a person to listen and form an impression, and impressions do not hold still. Two coaches score the same answer differently. The same coach scores it differently on a Friday.
Worse, listeners are measurably swayed by surface features of delivery that have nothing to do with what was said. Sixty years of perception research shows hesitation lowering ratings of speaker credibility, hedges and qualifiers costing perceived competence in a courtroom setting, and a 2025 study in PNAS finding that superficial auditory fluency shifts judgments of unrelated content. If the measuring instrument is a human ear, the measurement inherits every one of those biases.
A 1,000-point index, with every weight published
The Wellspoken Index is a 1,000-point scale built from six weighted dimensions and fifteen sub-metrics. We publish the complete methodology and the weight of every dimension at wellspoken.me/wellspoken-index/methodology, which is unusual for a proprietary score. If the weights are hidden, nobody can check whether the number means what we say it means.
| Dimension | Points | How it is scored |
|---|---|---|
| Structure | 250 | Sequencing, transitions, signposting, openings and closings |
| Conciseness | 200 | Word choice, plus an arithmetic formula for sentence length |
| Confidence | 150 | A multimodal model analyses the audio directly for hedging, uptalk and conviction |
| Pronunciation | 150 | Phoneme-level accuracy from a dedicated speech-assessment engine. Deterministic. |
| Fillers and Fluency | 150 | An exponential decay formula on fillers per minute. Deterministic. |
| Pace | 100 | A Gaussian curve centred on effective words per minute, with hesitation ratio. Deterministic. |
Structure and Conciseness together account for 450 of the 1,000 points, close to half. Most tools in this category lead with filler-word counts, which carry 150 points here, because filler words are the easiest thing in speech to count.
The research literature is far less certain than the advice industry that filler frequency alone is a defect. Filled pauses carry information for listeners about upcoming delay and speaker uncertainty, and their rate varies systematically with age, gender and personality. Counting them is useful. Treating the count as the whole story is not.
What having a score actually changes
A single number on its own would be a novelty. What makes it useful is what the decomposition supports.
It tells you where you are losing points, not that you are "unclear." Because the score decomposes into six dimensions and fifteen sub-metrics, you are not told to communicate better. You are told that Structure cost you 90 of 250 while Pace cost you nothing. That is an instruction rather than a verdict.
It separates problems that look identical from the outside. Two people who both sound nervous can be failing on completely different dimensions, one on hedging and inflection, the other on sequencing. Those need opposite exercises. Without a decomposed measure, both get the same generic advice about confidence, and one of them is wasting their time.
It tells you whether you actually improved. This is the part conventional coaching cannot supply, because a coach's impression of how you sounded last month is not a measurement. Scored practice produces a series, so progress becomes a line you can look at rather than a feeling you either have or do not.
Why this is not a chatbot with a voice on it
A general-purpose assistant reads a transcript and gives an opinion. We measure signals and return numbers. Those are different products even when the conversation feels similar.
With usable audio, 460 of the 1,000 points are computed from fixed scoring procedures rather than model judgment. Pronunciation (150) is scored per phoneme by a dedicated speech-assessment engine. Filler and fluency (150) runs on an exponential decay formula against fillers per minute. Pace (100) runs on a Gaussian curve around effective words per minute. Sentence length, 60 of the 200 Conciseness points, is arithmetic. A language model writes the coaching text for those dimensions and does not set the scores, so for a given input and scoring path those numbers repeat exactly. Pronunciation falls back to a text-only model estimate when audio or the speech engine is unavailable, so the deterministic share depends on the recording.
We listen to the audio, not only the transcript. Confidence scoring sends the actual recording to a multimodal model, so it can hear an upward inflection at the end of a sentence. A transcript cannot tell you that a statement ended like a question.
We remember you. Every ten sessions, a background process compresses your coaching history into a behavioural summary and feeds it into later sessions. A fresh chat window does not know you have said "kind of" forty times this month.
Score stability is an engineering problem we have to manage. When we changed the model behind one scoring tier, the same ten recordings moved by 43 to 70 points out of 1,000. We use replay testing against saved sessions to quantify that drift when scoring models change. A conversational assistant faces no such constraint, because nothing it says is a measurement.
Under the hood
A recording moves through a pipeline rather than a single prompt. Audio is transcribed with provider fallback, because recognisers differ substantially in whether they preserve filled pauses at all. A meeting recording is separated by speaker first, so only your voice is scored, which is what makes analysis of a real conversation possible. The selected speaker's segments are then filtered, padded and normalised, and the extracted audio is unavailable when usable speech totals under ten seconds, so the system does not score noise.
Six category evaluations run in parallel, followed by a seventh synthesis pass that writes the coaching summary.
Model selection is organised by latency tier rather than by vendor. The code requests a tier, never a named provider, and each tier holds a primary model with fallbacks across several providers. Real-time conversation runs on a separate stack built for low latency, with component and provider fallbacks behind it.
Within Index evaluation, failure degrades rather than stops: a single failed category returns a midpoint score, and if the whole Index fails the session is still saved.
What the data showed
Answer length turned out to be strongly bimodal. 52.2% of answers ran under 20 words, while 9.6% ran past 150 and 2.0% ran past 300. The mean of 53.6 words sits between the two clusters and matches few actual answers.
Within a typical session, the longest answer ran about six times the session median, so answer length varied substantially within a single conversation rather than settling at one habitual length. The study is descriptive and tests no hypothesis about why.
Responses to some version of tell me about yourself were isolated in 877 of the 1,000 sessions. The median ran 49 words, three times the median answer elsewhere in the corpus, and 14.6% ran past 150 words.
| Measure | 25th | Median | 75th | 90th |
|---|---|---|---|---|
| Words per answer | 5 | 17 | 72 | 147 |
| Answer to "tell me about yourself" | n/a | 49 | 103 | 188 |
The full study reports the complete distribution, the filled-pause rates, and seven limitations that we publish ahead of the results rather than behind them.
Built for the professional working in a second language
Wellspoken is built for people who are judged in a language they did not grow up speaking, where the gap between written fluency and spoken fluency is widest. You can write flawless business English and still lose control of a spoken answer, because writing allows revision and speaking does not.
The research here is unambiguous and uncomfortable. Accent affects judgments of credibility independently of whether the speaker was actually understood, and audit studies find accent shaping outcomes as early as the first screening call.
Phoneme-level scoring is what serves that user. Instead of being told your accent is strong, you are told which specific sound was unclear, which is an actionable instruction rather than a verdict. Systematic reviews of computer-assisted pronunciation training find that targeted, specific feedback is the component that actually moves learners.
The timing is not incidental. During 2026, as written application materials stopped carrying information about a candidate because anyone could generate them, unscripted speech became a hiring filter. One industry analysis reports applications reaching an average of 244 per open role, and survey work through 2026 tracks the same shift inside workplaces.
Practice, and real meetings
Beyond practice drills, our macOS desktop app records and analyses real calls on Google Meet, Zoom and Microsoft Teams, isolating your voice and scoring only you. Feedback comes from conversations that actually happened. Wellspoken runs on iOS, Android, macOS and in the browser.
The free tier includes three scored practice sessions per week and is permanent rather than a trial. The browser tools, including the filler-word counter and the speaking-rate checker, need no account at all.
