← SSI archive · Review rubric

2022 · arXiv / imported corpus page · Field expert review · confidence high

Multi-modality Associative Bridging through Memory: Speech Sound Recollected from Face Video

Minsu Kim, Joanna Hong, Se Jin Park, Yong Man Ro

The key idea is not generic fusion; it is storing cross-modal correspondences so video-only decoding can recover some audio-side structure later.

Verdict: full-text draftPriority: highConfidence: highBasis: full textCoverage: high

Reading guidance

Verdict
full-text draft · priority high · confidence high
Why it matters
The full text supports that the memory bridge is doing real work on benchmarks: gains are modest on LRW but large on LRW-1000 and enough to edge prior methods on GRID reconstruction.
What to trust
Basis: full text. Coverage: high. 4 evidence records back the review.
What is weak
The speech-reconstruction claim remains on speaker-dependent GRID, and the best subjective scores require an added WaveNet vocoder. No live deployment, no uncontrolled capture conditions, and no speaker-independent reconstruction result are shown. The method needs paired audio-video supervision and benchmark-like preprocessing. Benchmark lip reading and speech reconstruction from face video. Overclaim risk: The paper supports benchmark improvements, not a generally deployable memory-based SSI..
Read before
SSI review rubric
Read next
SSI archive

Axes

Task
speech-reconstruction
Modality
silent face video with recalled audio memory during training
Body site
face; lip
Output
speech-audio
Vocabulary
word-level lip reading plus fixed-phrase speech reconstruction
Metrics
word accuracy; STOI; ESTOI; PESQ; human naturalness and intelligibility ratings
Evaluation mode
word-level lip-reading benchmarks on LRW/LRW-1000 plus speaker-dependent GRID speech reconstruction with objective and human evaluation
Review confidence
high
Overclaim risk
The paper supports benchmark improvements, not a generally deployable memory-based SSI.

Expert take

Table 1 is the cleanest recognition evidence: the proposed method reaches 85.4 on LRW and 50.82 on LRW-1000, which is especially meaningful because the LRW-1000 gap over the next best method is large. The reconstruction side is smaller but still real: Table 2 reports 0.738 STOI, 0.579 ESTOI, and 1.984 PESQ on speaker-dependent GRID, edging Lip2Wav and Yadav et al. The human study in Table 3 reinforces that the gains are audible, with 2.93 naturalness and 3.56 intelligibility before WaveNet, and 4.37/4.27 with the WaveNet vocoder. The catch is scope: the reconstruction result is speaker-dependent GRID, not broad in-the-wild SSI.

True value

The full text supports that the memory bridge is doing real work on benchmarks: gains are modest on LRW but large on LRW-1000 and enough to edge prior methods on GRID reconstruction.

What changed

Canon before

Audio-visual fusion and common-representation methods usually needed both modalities at inference or failed when one modality was missing.

Delta from canon

This paper claims that memory can preserve cross-modal associations so a visual-only downstream model can still borrow audio structure.

Position in field

Interesting cross-modal architectural paper adjacent to lip reading and lip-to-speech reconstruction.

Evidence

“ That is, it can obtain both audio and visual tary information of different modalities and achieve high contexts during inference even when the uni-modal input is performance compared to the uni-modal methods. ”

author_claim · Abstract · confidence 0.99

“ Lip reading word accuracy comparison with visual modal word-level lip reading using only visual modal inputs on inputs on LRW and LRW-1000 dataset. benchmark datasets with the state-of-the-art methods. ”

metric · Table 1. Lip reading word accuracy comparison with visual modal inputs on LRW and LRW-1000 dataset. · confidence 0.99

“ Especially for LRW-1000, which is known are sampled with rate of 25fps and 16kHz, respectively. to be a difficult dataset due to unbalanced training samples, Following [50, 39], subjects 1, 2, 4, and 29 are taken for the proposed method attains a large improvement of 5.58% speaker-dependent task. ”

validation_scope · 4.2.1 Dataset · confidence 0.98

“ Performance of speech reconstruction comparison with Method STOI ESTOI PESQ visual modal inputs in a speaker-dependent setting on GRID. ”

metric · Table 2. Performance of speech reconstruction comparison with visual modal inputs in a speaker-dependent setting on GRID. · confidence 0.99

Limits

Technical limits

The speech-reconstruction claim remains on speaker-dependent GRID, and the best subjective scores require an added WaveNet vocoder.

Evaluation limits

No live deployment, no uncontrolled capture conditions, and no speaker-independent reconstruction result are shown.

Deployment limits

The method needs paired audio-video supervision and benchmark-like preprocessing.

Scope limits

Benchmark lip reading and speech reconstruction from face video.