First35 scores your delivery with a voice AI that goes beyond generic critique. Here is exactly what we measure, how, and how it becomes a single number you can chase.
Uploading a raw 30-second recording to ChatGPT or Gemini gets you a generic critique. First35 combines local audio measurements (pitch, volume, pauses, timing) with your personal practice trends and sends that rich context to our voice AI. Anyone can wire audio into an LLM. Few know what else to feed it.
Do you punch key words for impact? A flat, monotone-volume speaker scores low. A speaker who deliberately varies loudness to stress key words scores high.
Raw signal RMS amplitude per 50 ms frame across the recording.
Derived Dynamic range in dB. The difference between the 90th and 10th percentile of voiced-frame RMS (silence excluded).
Drag the slider to see how your dynamic range maps to a score.
Does your voice vary in pitch, or does it stay monotone? Wider pitch range sounds more engaged. But sing-song extremes also lose points. This is a Goldilocks curve.
Raw signal Fundamental frequency (F0) per frame from autocorrelation.
Derived Interquartile range of F0, divided by median F0 to normalize male/female voices.
Ideal band 0.10 to 0.40. Below is monotone, above is sing-song.
Do your sentences end with confidence, or trail up like questions? Rising terminals on statements, called uptalk, make you sound unsure of your own words. Listeners hear it before they hear the words: the person delivering this doesn't fully believe what they're saying.
In interviews, pitches, and networking, uptalk is one of the most reliable predictors of being perceived as less confident, less experienced, and less senior, regardless of actual credentials. Fixing it is often the single fastest way to sound more authoritative.
Raw signal Per-frame F0 (same pitch track from Metric 2).
Derived Linear regression slope over the last 400 ms of each phrase, in semitones/sec. Phrases are segmented at silences ≥ 400 ms.
Slope > +2 st/s = rising = uptalk. Score = 100 × (1 − uptalk_phrases / total_phrases).
Each dot is one phrase. Green endings (falling/flat) are confident; red endings (rising) are uptalk.
Do you speak with a natural rhythm, or does it break into awkward silences? Deliberate short pauses help. Many long pauses signal hesitation.
Raw signal Silence mask: frames where RMS < −45 dBFS.
Derived Count and duration of silence runs ≥ 400 ms.
2 pauses = free. Beyond that, −10 each. Long pauses (> 2 s) add extra penalty.
The overall score is the plain average of the four sub-scores. Equal weight, no magic.
Try adjusting each metric to see how it moves the overall score:
Equal weighting is a deliberate choice today. The four metrics measure genuinely independent qualities. Nervous but articulate speakers score high on Pause/flow and low on Emphasis; confident but monotone speakers score high on Uptalk and low on Expressiveness. Equal weight rewards balanced delivery.
Future versions will calibrate the weights per region. Different parts of the world perceive each of these qualities differently, and what sounds confident in one culture can read as aggressive in another. As the app learns how listeners in each market actually respond, the mix will adapt.
Words-per-minute is measured using speech recognition and shown on the feedback screen as raw information. It is excluded from the composite because transcription is not always reliable. Including it would risk injecting noise into the overall score for reasons unrelated to your actual delivery.