Multimodal AI Hackathon 2026 · Track C

Speech Delivery Coach

Contrastive speech analytics & temporal flaw grounding. It finds where a delivery deviates from a reference reading of the same text, explains why in numbers, and coaches the fix. Everything below was produced by the real system.

▶ Run it live (Colab)Code · dataset · docs on GitHubTechnical documentation (PDF)

Demo video

Try the analyses

Three real runs on files from the contrastive dataset. Hover heatmap words, zoom the timeline, play A/B clips.

Reference: Ronald Reagan — Address on the Space Shuttle Challenger, 28 Jan 1986
Reading: Botched reading: four hidden flaws · · PDF report ↓

6.1/10 Overall delivery
6.2Pacesyllable rate stays close to the reference delivery
2.0Pausingno unplanned silences
4.8Expressivenesspitch keeps moving to carry emphasis
10.0Volume & clarityloud, bright and well-articulated
7.4Fluencyno repeated onsets or sounds outside the text

Practise these first

  1. Unplanned long pause at 13.5–16.5 s · Pausing, ~1.6 pts
    Close the gap at “and … I”: about 0.18 s is enough here (you paused 2.98 s). Mark the next phrase in your script so the pause is a choice, not a search.
  2. Monotone delivery at 22.6–24.9 s · Expressiveness, ~1.0 pts
    Give “our country. This is truly a” a melody: lift your pitch on “This” by about 1 semitone and let it fall at the end of the phrase (the reference moves 1.9 st here, you 0.5).
  3. Rushed pace at 4.0–4.9 s · Pace, ~0.8 pts
    Slow down on “the state of the Union, but”: aim for about 6.3 syllables/s (you were at 9.2). Take a breath before the phrase and land each stressed syllable.

Your style vs the reference

tempo +0%pauses ×1.00pitch range ×1.00voice 102 Hz vs 103 Hz
  • Overall tempo, pauses and pitch range are close to the reference.

Local flaws are measured against your own style, so a naturally slower or lower voice is not penalized everywhere.

Word heatmap

Ladies and gentlemen, I'd planned to speak to you tonight to report on the state of the Union, but the events of earlier today have led me to change those plans. Today is a day for mourning and remembering. Nancy and I are pained to the core by the tragedy of the shuttle Challenger. We know we share this pain with all of the people of our country. This is truly a national loss. Nineteen years ago, almost to the day, we lost three astronauts in a terrible accident on the ground. But we've never lost an astronaut in flight; we've never had a tragedy like this.

Each word is colored by its largest deviation from the reference (pace, pause, pitch, volume, articulation). Hover a word for its numbers.

Timeline: reference time-warped onto the reading

Ground truth (injected, hidden from the detector): rushed sev 3 at 3.98–4.94 s; long_pause sev 3 at 13.56–16.36 s; monotone sev 3 at 22.60–25.52 s; stutter sev 3 at 33.92–35.18 s

Flaws: numbers, what to do, and A/B listening

#flawwhy (numbers)what to dolisten
#1
3.96–4.94 s		 Rushed pace
“the state of the Union, but”		 Speech rate 9.2 syl/s vs baseline 6.2 syl/s (+47%, z=+11.2). Words are compressed; listeners lose the content.Slow down on “the state of the Union, but”: aim for about 6.3 syllables/s (you were at 9.2). Take a breath before the phrase and land each stressed syllable. A · reference B · reading
#2
13.47–16.45 s		 Unplanned long pause
“and … I”		 Silence of 2.98 s vs 0.18 s in the baseline (+2.80 s, z=+23.3) between "and" and "I". A pause this long reads as a lost line, not emphasis.Close the gap at “and … I”: about 0.18 s is enough here (you paused 2.98 s). Mark the next phrase in your script so the pause is a choice, not a search. A · reference B · reading
#3
22.63–24.95 s		 Monotone delivery
“our country. This is truly a”		 Pitch movement 0.5 st vs baseline 1.9 st (-74%, z=-11.2). Intonation is flat, so key words get no emphasis.Give “our country. This is truly a” a melody: lift your pitch on “This” by about 1 semitone and let it fall at the end of the phrase (the reference moves 1.9 st here, you 0.5). A · reference B · reading
#4
33.88–34.84 s		 Disfluency (repeat/stutter)
“we've”		 0.70 s of voiced sound before "we've" that is not in the transcript (0.40 s more than the baseline at the same word). Typical of a repeated onset or a filler.Restart cleanly on “we've”: if the word catches, stop for a beat and say it whole. Practise the phrase slowly three times, then at full speed. A · reference B · reading

Reference: John F. Kennedy — Address at Rice University on the Nation's Space Effort, 12 Sep 1962 (closing)
Reading: Near-perfect reading: two subtle flaws · · PDF report ↓

9.3/10 Overall delivery
10.0Pacesyllable rate stays close to the reference delivery
10.0Pausingno unplanned silences
6.4Expressivenesspitch keeps moving to carry emphasis
10.0Volume & clarityloud, bright and well-articulated
10.0Fluencyno repeated onsets or sounds outside the text

Practise these first

  1. Monotone delivery at 18.4–19.4 s · Expressiveness, ~0.7 pts
    Give “for knowledge and peace” a melody: lift your pitch on “peace” by about 2 semitones and let it fall at the end of the phrase (the reference moves 3.2 st here, you 1.3).

Your style vs the reference

tempo +0%pauses ×1.00pitch range ×1.00voice 176 Hz vs 174 Hz
  • Overall tempo, pauses and pitch range are close to the reference.

Local flaws are measured against your own style, so a naturally slower or lower voice is not penalized everywhere.

Word heatmap

Many years ago the great British explorer George Mallory, who was to die on Mount Everest, was asked why did he want to climb it. He said, because it is there. Well, space is there, and we're going to climb it, and the moon and the planets are there, and new hopes for knowledge and peace are there. And, therefore, as we set sail we ask God's blessing on the most hazardous and dangerous and greatest adventure on which man has ever embarked.

Each word is colored by its largest deviation from the reference (pace, pause, pitch, volume, articulation). Hover a word for its numbers.

Timeline: reference time-warped onto the reading

Ground truth (injected, hidden from the detector): rushed sev 1 at 4.81–7.20 s; monotone sev 1 at 17.99–20.14 s

Flaws: numbers, what to do, and A/B listening

#flawwhy (numbers)what to dolisten
#1
18.42–19.44 s		 Monotone delivery
“for knowledge and peace”		 Pitch movement 1.3 st vs baseline 3.2 st (-60%, z=-13.1). Intonation is flat, so key words get no emphasis.Give “for knowledge and peace” a melody: lift your pitch on “peace” by about 2 semitones and let it fall at the end of the phrase (the reference moves 3.2 st here, you 1.3). A · reference B · reading

Reference: Franklin D. Roosevelt — Day of Infamy address to Congress, 8 Dec 1941
Reading: One injected pause (severity 3) · · PDF report ↓

8.4/10 Overall delivery
10.0Pacesyllable rate stays close to the reference delivery
2.1Pausingno unplanned silences
10.0Expressivenesspitch keeps moving to carry emphasis
10.0Volume & clarityloud, bright and well-articulated
10.0Fluencyno repeated onsets or sounds outside the text

Practise these first

  1. Unplanned long pause at 37.8–40.7 s · Pausing, ~1.6 pts
    Close the gap at “States … was”: about 0.06 s is enough here (you paused 2.86 s). Mark the next phrase in your script so the pause is a choice, not a search.

Your style vs the reference

tempo +0%pauses ×1.00pitch range ×1.00voice 191 Hz vs 191 Hz
  • Overall tempo, pauses and pitch range are close to the reference.

Local flaws are measured against your own style, so a naturally slower or lower voice is not penalized everywhere.

Word heatmap

Mr. Vice President, Mr. Speaker, Members of the Senate, and of the House of Representatives: Yesterday, December seventh, nineteen forty one, a date which will live in infamy, the United States of America was suddenly and deliberately attacked by naval and air forces of the Empire of Japan. The United States was at peace with that Nation and, at the solicitation of Japan, was still in conversation with its Government and its Emperor looking toward the maintenance of peace in the Pacific.

Each word is colored by its largest deviation from the reference (pace, pause, pitch, volume, articulation). Hover a word for its numbers.

Timeline: reference time-warped onto the reading

Ground truth (injected, hidden from the detector): long_pause sev 3 at 37.84–40.64 s

Flaws: numbers, what to do, and A/B listening

#flawwhy (numbers)what to dolisten
#1
37.81–40.66 s		 Unplanned long pause
“States … was”		 Silence of 2.86 s vs 0.06 s in the baseline (+2.80 s, z=+23.3) between "States" and "was". A pause this long reads as a lost line, not emphasis.Close the gap at “States … was”: about 0.06 s is enough here (you paused 2.86 s). Mark the next phrase in your script so the pause is a choice, not a search. A · reference B · reading

Stress test: labels hidden from the detector

81%of injected flaws found
0false alarms on clean speech
10 → 9.0 → 5.3score: clean → near-perfect → botched
4/9reversed words found as unclear

Detection by flaw and severity

flawsev 1sev 2sev 3start error (median)
dragging3/33/32/30.078 s
long pause3/33/35/60.032 s
monotone5/63/36/60.335 s
mumble3/33/33/30.119 s
rushed3/63/36/60.051 s
stutter0/31/33/60.045 s

Weakest: single stutters and articulation. Thresholds were tuned on this dataset, so these numbers are optimistic; a held-out speaker is next.

How it works

Three public-domain speeches (FDR 1941, JFK 1962, Reagan 1986) are mirrored by 6 flaw types × 3 severities, injected by signal processing so every timestamp is exact: 63 labeled files. Both readings are force-aligned to the same transcript (Meta MMS), and a single word-level delta table (pitch in semitones, energy in dB, articulation rate, pauses, MFCC shape) drives the style profile, the flaw regions, the rubric, the tips, the heatmap, A/B clips and the PDF report.