diddmstjr/confusable-100-lipreading

Confusable-100: an English vocabulary built to break lip reading

A small, deliberately adversarial corpus for probing one specific failure mode of visual speech recognition: consonants articulated by the tongue leave no distinctive trace on the lips, so words differing only in those consonants are not separable from video in principle, not merely in practice.

This is an evaluation probe, not a training set. It is one speaker and 3.4 minutes of speech. Its purpose is to expose structure in the errors a recognizer makes, not to measure competitive accuracy.

Why it exists

No public corpus is deliberately concentrated on lingual minimal pairs, so we designed a vocabulary and recorded it.

The 101-word vocabulary holds the visible portion of each word as nearly constant as possible while varying the invisible portion. It is dominated by monosyllables of the form C3–V–C6: a maximally salient bilabial onset, one of four vowel classes, and one or two maximally occluded lingual codas.

Under a standard phoneme-to-viseme mapping the vocabulary collapses onto 45 distinct viseme sequences, and 73 of 101 words (72.3%) are homophenous with at least one other — 2.4× a length-matched random English control (30.5%).

A finding worth stating: lingual phonemes are 29.7% of our phoneme tokens against 36.7% in the control. The vocabulary is not confusable because it contains more tongue-articulated segments. The collapse is structural, not compositional — it comes from minimal pairs whose only contrast is carried by a hidden segment.

back / bad / ban / bat differ only in the final consonant — /k, d, n, t/ — all articulated by the tongue inside the closed mouth.

Contents

100 recordings (100 of the 101 word types), one utterance per word, one speaker, one session:

{index}_{word}/
  video.mp4        mouth-region video
  emg.npy          (3, T) float32   see the warning below
  metadata.json    text, duration, fps, frame count, EMG rate, timestamp

• Video: nominal 30 fps (measured 30.2–30.9, mean 30.5)
• Utterances: mean 2.03 s (1.55–2.63), mean 61 frames (46–79); 3.4 min total
• Three sEMG channels recorded simultaneously at mean 251.9 Hz

⚠️ The EMG channels contain no signal — this is deliberate documentation

The sEMG in this corpus carried no myoelectric content and must not be used to train or evaluate anything. It is included because the negative result is part of the scientific record, and because it is a useful teaching case.

Four diagnostics over 98 of the 100 utterances:

Diagnostic	Genuine sEMG	Observed here
Speech vs. silence, 20–120 Hz power	+6 to +20 dB	−2.3 / −1.5 / −1.3 dB
Spectral shape	peak 50–150 Hz	flat to the 127 Hz Nyquist limit
60 Hz mains peak	present	±0.6 dB (absent)
Amplifier output	above quantization floor	12–50 ADC counts

The band that should gain power during articulation instead lost it; the spectrum is that of a stationary noise process; and the complete absence of mains pickup is most consistent with poor electrode–skin coupling. The fourth row only rules out an inert amplifier: the instrument was digitizing its own noise.

This matters beyond this corpus. A modality carrying no information yields a null fusion gain that is easily misread as an architectural failure. Verify that signal exists before attributing a null result to your model.

The video channel is valid and is what the companion work uses.

Reference results

Decoding all 100 words zero-shot with a pretrained visual-only Auto-AVSR checkpoint (beam search, no fine-tuning, no external language model):

Quantity	Value
Word accuracy	0.19
Mean phoneme error rate	0.561
Spearman ρ, confusability vs. per-word PER	0.223
Permutation p	0.024
Permutation p, word length controlled	0.045

The association is weak — about 5% of rank variance — and should be read as supportive rather than decisive. Errors are structured: among consonant viseme classes the lingual class has the highest error rate (0.505), and word-final consonants are recovered less often than word-initial ones (0.634 vs 0.470). The effect is directional but not surgical: visible onsets fail on nearly half the words too, because a wholly mistaken word hypothesis corrupts every position at once.

Limitations

• One speaker, one session, 3.4 minutes. Absolute error rates will not generalize. Conclusions concern the structure of errors, not their level.
• Adversarial by construction. The collapse rate is not an estimate of the collapse rate of English; it isolates one failure mode against a controlled baseline.
• The EMG is unusable (see above).
• Not suitable for training. Used only for zero-shot evaluation in the companion work.

Human subjects

This corpus is video of a single, identifiable speaker — the author, who recorded and released it voluntarily. It contains mouth-region video and no other identifying metadata. If you use it, please respect that it depicts a real person: use it for research on visual speech recognition, and do not use it to build or evaluate biometric identification systems.

License

CC-BY-NC-SA-4.0. Non-commercial use only; attribution required; derivatives must carry the same license.