← SSI archive · Review rubric

2023 · arXiv / imported corpus page · Field expert review · confidence high

Intelligible Lip-to-Speech Synthesis with Speech Units

Jeongsoo Choi, Minsu Kim, Yong Man Ro

Speech units as a pseudo-text target enable strong content supervision that substantially cuts WER without text labels, and the multi-input vocoder improves speech quality from blurry mel outputs, yielding a state-of-the-art lip-to-speech system on LRS benchmarks.

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

Reading guidance

Verdict
full-text draft · priority high · confidence high
Why it matters
A strong contribution to video-based speech reconstruction that eliminates the need for paired text transcriptions by leveraging self-supervised speech units, enabling intelligible speech synthesis with improved content fidelity. The multi-input vocoder innovation also materially enhances waveform quality from predicted features.
What to trust
Basis: full text. Coverage: high. 6 evidence records back the review.
What is weak
Limited by dependence on benchmark video quality, vocoder stack complexity, and residual challenge in matching natural speech quality; no tested robustness to pose, occlusion, or spontaneous environment variability. Evaluation is limited to test sets from LRS2 and LRS3 datasets, with no cross-dataset or live deployment exploration; human evaluation is limited to 15 participants rating MOS on 20 samples. No latency analysis, on-device implementation, or in-the-wild robustness to camera position/lighting/occlusion is demonstrated, limiting immediate real-world deployment. Focuses strictly on lip-to-speech synthesis (speech reconstruction) without addressing text entry, command recognition, or speech enhancement. Evaluation confined to standard benchmarks LRS2 and LRS3. Overclaim risk: medium-low.
Read before
SSI review rubric
Read next
SSI archive

Axes

Task
speech-reconstruction
Modality
video
Hardware
Lip video with visual front-end optionally replaced by AV-HuBERT models pretrained on LRS3 and VoxCeleb2
Body site
lip
Output
speech-audio
Metrics
Metrics used include STOI, ESTOI, PESQ, and Word Error Rate (WER) on LRS2 and LRS3 datasets, plus Mean Opinion Score (MOS) for naturalness, intelligibility, and clearness based on human raters.
Evaluation mode
Quantitative evaluation on standard benchmarks (LRS2, LRS3) using STOI, ESTOI, PESQ, WER metrics and a small-scale human Mean Opinion Score (MOS) study.
Review confidence
high
Overclaim risk
medium-low

Expert take

This paper presents a significant advancement in lip-to-speech synthesis by combining self-supervised quantized speech units as auxiliary prediction targets with a multi-input vocoder that conditions on both mel-spectrograms and speech units. This architecture allows improved speech content modeling without text labels, a key bottleneck in prior work. Experimental results on LRS3 and LRS2 benchmarks demonstrate clear metric improvements, notably a reduction in word error rate (WER) from 65.8% to 29.8% when using a strong AV-HuBERT visual encoder plus speech units. The vocoder design, augmented during training with blur and noise, further enables intelligible speech synthesis from blurry predicted mel features. Human evaluations confirm gains in naturalness, intelligibility, and clearness relative to prior state of the art. Despite these merits, the work remains benchmark-focused without investigation of real-time operation, mobile compatibility, or robustness to real-world video variations such as occlusion or pose changes. The approach significantly moves the field towards practical lip-to-speech reconstruction by eliminating reliance on text labels, but future work must address deployment and generalization challenges.

True value

A strong contribution to video-based speech reconstruction that eliminates the need for paired text transcriptions by leveraging self-supervised speech units, enabling intelligible speech synthesis with improved content fidelity. The multi-input vocoder innovation also materially enhances waveform quality from predicted features.

What changed

Canon before

Lip-to-speech models typically either use blurry acoustic targets alone for supervision or require paired text labels as stronger content guidance.

Delta from canon

Uses discrete speech units derived via speech model quantization as pseudo-text supervision to the multi-target L2S model, enabling improved content modeling without text labels; additionally, uses speech units to condition a multi-input vocoder for waveform generation, enhancing intelligibility and speech quality.

Position in field

Core video-to-speech / silent-speech reconstruction work utilizing self-supervised speech units for content modeling.

Evidence

“ This task can re- eling content in the output speech without any additional labels. solve the limitations of lip-reading by training an L2S model on In addition, we propose a novel multi-input vocoder, which audio-visual data which are more easily available than video- is for converting the synthesized mel-spectrogram into wave- text paired data. ”

author_claim · Abstract · confidence 1.00

“ Moreover, to reduce the gap between the vision to the L2S model even without using additional labels. train and test data, we augment the input mel-spectrogram with The speech units can be obtained by quantizing speech repre- blur and noise during training. ”

actual_novelty · 2.2. Multi · confidence 1.00

“ Table 3: Ablation Study on LRS3 Dataset the predicted speech units from the multi-target L2S model are Method STOI ESTOI PESQ WER(%) not perfect, they can serve as an additional discrete condition Baseline [8] 0.516 0.292 1.27 72.5 for the vocoder to reduce artifacts and help the model generate + Speech units 0.542 0.343 1.29 50.7 high-fidelity waveforms. ”

metric · 4.1 Quantitative Comparison · confidence 1.00

“ A total of 15 participants are asked Table 1 and 2 show the evaluation results of the proposed meth- to rate their opinions on naturalness, intelligibility, and clear- ods and the previous methods on LRS3 and LRS2, respectively. ness on a scale of 1 (least) to 5 (most). ”

validation_scope · 3.1 Datasets · confidence 1.00

“ Table 3: Ablation Study on LRS3 Dataset the predicted speech units from the multi-target L2S model are Method STOI ESTOI PESQ WER(%) not perfect, they can serve as an additional discrete condition Baseline [8] 0.516 0.292 1.27 72.5 for the vocoder to reduce artifacts and help the model generate + Speech units 0.542 0.343 1.29 50.7 high-fidelity waveforms. ”

limitation · 4 Results · confidence 1.00

“ While the paralinguistic information and mainly keeps the linguistic the performance of lip-reading models has improved signifi- information. ”

deployment_claim · 5 Conclusions · confidence 1.00

Limits

Technical limits

Limited by dependence on benchmark video quality, vocoder stack complexity, and residual challenge in matching natural speech quality; no tested robustness to pose, occlusion, or spontaneous environment variability.

Evaluation limits

Evaluation is limited to test sets from LRS2 and LRS3 datasets, with no cross-dataset or live deployment exploration; human evaluation is limited to 15 participants rating MOS on 20 samples.

Deployment limits

No latency analysis, on-device implementation, or in-the-wild robustness to camera position/lighting/occlusion is demonstrated, limiting immediate real-world deployment.

Scope limits

Focuses strictly on lip-to-speech synthesis (speech reconstruction) without addressing text entry, command recognition, or speech enhancement. Evaluation confined to standard benchmarks LRS2 and LRS3.