Get to know your coach while it works. Preparation and analysis continue as you read here.
Frequently asked questions
Which coaching rules does Speech Coach use?
These are the 8 checks. They cover vocal variety, pace, pauses, gestures and purposeful movement.
1. Filler words and sounds
The Fillers tab counts transcript matches. English: um, uh, erm, er, ah, hmm, umm, uhh, you know, I mean, kind of and sort of. German: äh, ähm, öhm, hm and hmm. Three or more matches in a 30-second window create a review moment. Try replacing an unnecessary filler with a pause. Transcription can omit fillers; a matched phrase may also carry meaning, so listen back before changing it.
2. Vocal variety — pitch, volume and pauses
The Voice tab checks 12-second sections for both low pitch variation (standard deviation below 7.5% of mean pitch) and low volume variation (standard deviation below 2.6 dB). Both must be low to flag a section. Quiet intervals of at least 0.35 seconds count as pauses. The voice snapshot describes pause frequency and length, pitch variation and volume changes. These measurements suggest where to practise emphasis; they do not judge emotional tone.
3. Pace — speaking rate
The Pace tab estimates words per minute in roughly 30-second windows. A final remainder shorter than 15 seconds joins the previous window. It needs at least 20 seconds of timed speech and 30 words. The practice range is 120–160 words per minute; internal pauses count, leading and trailing silence do not. Faster or slower passages become review suggestions. Adapt the pace to the language, audience and purpose.
4. Gestures and purposeful movement — body language
The Body language tab flags hands visible in fewer than 45% of video samples. In 20-second windows it checks for gesture speed below 45% of your recording average or gesture size above 1.8 times your average. It also flags pronounced side-to-side movement. These are prompts to review framing and purposeful movement, not instructions to keep moving or stay perfectly still.
5. Facial expressions — supporting the message
The Expression tab matches positive, difficult or surprising wording with nearby visible facial activity. It flags little facial change around strongly positive or surprising words, or a strong smile around negative wording. Review whether the contrast is intentional. The app does not infer your feelings or assess eye contact.
6. Emphasis — matching delivery to meaning
The Message & delivery tab finds wording cues for lists, contrasts, important points, change, size, emotion, conclusions and questions. It checks for nearby gesture support, vocal contrast, facial change or a pause of at least 0.45 seconds, depending on the cue. Two or more missing supports create a review moment. A purposeful pause or gesture may help; every cue does not require every kind of emphasis.
7. Word choice and grammar
The Language tab checks immediate word repetitions and transcript segments of 45 words or more. English rules suggest because for due to the fact that, to for in order to, now for at this point in time, and reducing very very. They flag basically, actually and literally for review, plus simple don't/doesn't agreement errors. German rules suggest weil for aufgrund der Tatsache, dass, jetzt for zum jetzigen Zeitpunkt, können for in der Lage sein, and reducing sehr sehr. This is a small rule set, not a full grammar or pronunciation assessment.
8. Descriptive language — metaphors and comparisons
The Metaphor ideas tab matches wording about complexity, growth, foundations, risk, transformation, collaboration and barriers with a fixed library of images. It checks for example or comparison cues such as like a, imagine and for example, or wie ein, stell dir vor and zum Beispiel. It does not judge originality or whether a metaphor is appropriate; choose an image only when it helps your audience.
How are scores and review moments chosen?
Most radar areas need at least 30 transcript words plus the relevant voice or video evidence. Missing evidence stays unscored. Message & delivery, Expression, Body language, Voice and Language use the share of measured checks, time or passages without the corresponding flags. Fillers scores 25 with no matches and falls to 0 at 10 matches per 100 words. Pace scores 25 within 120–160 words per minute and declines outside that band. Metaphor ideas uses the share of matched abstract passages containing an example or comparison cue. Open How the scores work below the radar for each formula.
The review list shows up to ten moments. Its fixed priority order is filler clusters, missed emphasis, facial-expression contrasts, low vocal variety, body-language observations, pace, wording and metaphor ideas. Similar moments of the same type within four seconds are merged. The score is a practice prompt, not a competition mark or an assessment of confidence, credibility, speech structure or audience engagement.
How reliable is the remaining-time estimate?
The rough estimate uses the progress and processing speed observed on this device. It recalculates when the analysis stage changes and may increase or disappear while a long step is running. Keep Speech Coach open: switching tabs, locking the screen or power saving can slow or pause processing. If notifications are enabled and your browser has granted permission, we will notify you when the report is ready, including if it finishes earlier.
How can I send feedback to DeepMirror?
Use the speech bubble beside the ? button. It opens DeepMirror's existing feedback form in a new tab, keeping this analysis open. Sign in if prompted. Feedback goes to support@deepmirror.me; your recording and report are not attached automatically.
Does this use ChatGPT, Claude or similar generative AI?
No. Your recording is not sent to ChatGPT, Claude, Gemini or another cloud chatbot. Speech Coach uses small speech and vision models directly on your device, then applies fixed coaching rules to the results. This keeps your recording private and avoids sending your speech to a generative-AI service.
Does my recording leave my device?
No recording upload is needed for analysis. For the strongest privacy check, wait until setup shows “Ready”, then switch off Wi‑Fi or use Airplane Mode before choosing the recording.
Who created Speech Coach?
Speech Coach was created by the DeepMirror.me team as part of our work on privacy-first tools that help people reflect, learn and communicate better.
Why did you build this?
We want practical self-development tools to be more accessible without asking people to pay monthly subscriptions or give up sensitive personal data. Speech Coach is being developed as a public-benefit DeepMirror.me tool focused on useful, private feedback rather than advertising or data monetization.
Which recordings can I use?
You can choose video or audio from iPhone, Android, Mac or PC. Common formats such as MOV, MP4, M4A, MP3 and WAV are accepted. Analysis starts automatically when the recording can be read on your device. If a particular codec cannot be read, Speech Coach tells you rather than uploading or converting it elsewhere.
Can I leave the app in the background?
You can read this Help view while preparation and analysis continue. Use the ? button or “Back to Speech Coach” to return without losing your place or recording. The progress summary above stays up to date, and shows “View report” when your feedback is ready.
Switching to another browser tab or locking your phone is different: your browser may slow or suspend processing. Keep Speech Coach open for the most reliable result. You can enable “Notify me when done” in Options.
Why does it take a little time to get ready?
Speech Coach loads its speech, video and audio tools onto your device before you choose a recording. It also checks that the local media decoder is actually working before showing Ready.
Technical details
Speech recognition and body/face analysis run in your browser using local machine-learning models. For transcription, Speech Coach uses longer overlapping listening windows so words near a boundary keep their context. The coaching report combines transcript, voice, pose, hand and facial signals using deterministic rules. No cloud generative language model is used for the analysis or report.