Source-linked AI summary
CAPQ-FAST: Content-Adaptive Perceived Quality Assessment for Faster Audiovisual Playback
Jiarun Song, Yuxin Song, Fuzheng Yang, Weisi Lin
TL;DR
The paper addresses limited evidence on how playback speed affects perceived quality across different media modalities and content characteristics. It conducts subjective experiments, extracts modality-specific temporal features, and integrates separate predictive models into CAPQ-FAST. The resulting model predicts perceived quality across playback speeds and supports interpretation of faster-playback intentions, while remaining limited by dataset, content, and participant diversity.
Problem
Systematic evidence and objective methods remain limited for perceived quality under faster playback across video, audio, and audiovisual modalities.
Method
Subjective experiments measure perceived quality across modalities, while TI, WPM, and BPM support separate video, speech, and music models integrated into CAPQ-FAST.
Results
CAPQ-FAST effectively predicts perceived quality under different playback speeds and outperforms baseline models against subjective ratings.
Takeaways & Limitations
Perceived quality provides an auxiliary cue for interpreting faster-playback intentions and informing personalized recommendations and playback control.
Takeaways & Limitations
The dataset, content diversity, and participant diversity are limited, while TI, WPM, and BPM do not capture semantic importance, narrative relevance, or user interest.
Abstract
from arXiv · showhide
Faster playback has become a common feature in modern online audiovisual services, allowing users to consume content in less time while still maintaining a coherent viewing experience. However, different modalities of media content, such as video, audio (including speech and music), and audiovisual, exhibit varying requirements for understandability and information integrity under faster playback. These differences lead to noticeable variations in perceived quality depending on the content type. Nevertheless, users' perceived quality at different playback speeds remains insufficiently investigated. To address this gap, this paper conducts a series of subjective experiments to analyze the relationship between playback speed and perceived quality for video, audio, and audiovisual content. Content-specific intrinsic features are extracted to capture temporal dynamics, including temporal information (TI) for video, words per minute (WPM) for speech, and beats per minute (BPM) for music. Predictive models of perceived quality under faster playback are then developed separately for video, speech, and music. By integrating these models, a unified content-adaptive perceived quality assessment model (CAPQ-FAST) is proposed for faster audiovisual playback. Experimental results demonstrate that the proposed model effectively predicts perceived quality under different playback speeds. This model can help service providers better understand users' viewing intentions and perceptual experiences under faster playback, thereby enabling more adaptive personalized recommendations and playback control to enhance user experience adaptability.
I. INTRODUCTION
Faster playback supports efficient media consumption but creates ambiguous signals about user intent because perceived quality depends on content modality and temporal characteristics. The paper addresses this gap by developing CAPQ-FAST from modality-specific features and perceived-quality models.
- Faster playback helps users consume audiovisual content more efficiently and reflects increased flexibility and autonomy in media consumption.
- The same playback speed can indicate improving information efficiency or skipping irrelevant content, making speed alone insufficient for inferring user intent.
- Perceived quality at variable speeds can provide an auxiliary basis for recommendation strategies and personalized playback control.
- Video, audio, and audiovisual content differ in understandability and information-integrity requirements under faster playback, producing modality-specific perceived-quality changes.
- CAPQ-FAST integrates playback speed with temporal information (TI), words per minute (WPM), and beats per minute (BPM) to assess perceived quality adaptively.
II. RELATED WORK
Prior research links faster playback to cognitive load, comprehension, and playback optimization, but it largely emphasizes educational lecture videos and system performance. This paper extends the focus toward diverse real-world content and user perception.
- Cognitive load of faster playback: Cognitive-load research reports impaired comprehension when information presentation exceeds processing capacity, with effects often pronounced beyond approximately twice normal speed.
- Cognitive load of faster playback: Educational studies find that moderate speed increases can improve time efficiency, while other evidence reports declining comprehension or no significant impairment within specific thresholds.
- Research scope: Most prior studies focus on educational lecture videos with simple structures and speech-dominant information delivery, limiting coverage of real-world online media.
- Playback optimization: Faster playback compresses temporal presentation and increases information delivered per unit time, unlike reduced-frame-rate playback, which preserves duration but introduces discontinuity.
- Playback optimization: Adaptive playback technologies mainly optimize latency and bitrate while often overlooking users’ perceptions and differing motivations for selecting faster speeds.
III. DESIGN OF EXPERIMENTS
The study uses a dedicated subjective-testing platform to collect perceived-quality ratings for video, audio, and audiovisual content at multiple playback speeds. The platform supports controlled playback and automated rating capture.
- Subjective tests collected perceived-quality ratings across video, audio, and audiovisual content at different playback speeds for modeling and validation.
- An Android-based platform supports faster playback from 1x to 3x and reduces repeated manual speed adjustments between media sequences.
- The platform automatically records ratings and presents content on smartphones with 5.9-inch screens and headphones.
B. Participants
The experiment included 72 university-student participants aged 22–28, with normal or corrected-to-normal vision and hearing and substantial familiarity with faster playback.
- 72 university students participated, with ages ranging from 22 to 28 years (M = 24.2, SD = 2.3).
- The sample included 33 females (45.8%) and 39 males (54.2%), and all participants had normal or corrected-to-normal vision and hearing.
- Most participants had prior faster-playback experience, supporting stable and natural judgments under different playback speeds.
C. Experimental Settings and Procedures
The study defines perceived quality as users’ judgment of whether content remains clear, followable, and acceptable under faster playback, then evaluates video, audio, and audiovisual material through subjective tests. The procedures vary content characteristics and playback speeds to support later modeling and validation.
- Perceived quality definition: Perceived quality covers clarity, understandability, followability, and acceptance, with modality-specific emphasis for video, speech, and music.Video and speech emphasize clarity and understandability, while music emphasizes auditory acceptability.
- Experimental design: The study conducts subjective evaluations across video, audio, and audiovisual content to examine perceived quality at different playback speeds.The experiments provide ratings for perceptual analysis, model construction, and validation.
- Platform comparison: The preliminary comparison tests the developed platform against YouTube, TikTok, Kwai, VLC, and KMPlayer at five playback speeds across three media modalities.Each modality uses three test sequences, with pairwise perceived-quality comparisons at 1x, 1.25x, 1.5x, 2x, and 3x.
- Video session: The video training experiment uses 75 high-quality 1080p lossless sequences spanning motion complexities and content categories, tested at 1x, 1.25x, 1.5x, 2x, and 3x.The sequences include news, movies, dance, sports, and entertainment videos.
- Audio session: Audio testing separates speech and music, grouping speech by words per minute and music by beats per minute before evaluating five playback speeds with five-point ACR ratings.Speech includes narration, lectures, news, and conversation; music covers different styles and rhythmic characteristics.
- Audiovisual and validation sessions: Audiovisual testing uses speech-video and music-video sequences at five playback speeds, while validation includes passive viewing for prediction and active playback for modeling viewing intentions.The validation design distinguishes fixed-speed evaluation from participants’ freely selected playback speeds.
IV. EXPERIMENTAL RESULTS FOR MODELING
The modeling section frames CAPQ-FAST as a unified assessment built from modality-specific perceived-quality models and content features. Its stated pipeline compares platform performance, models video and audio quality, and analyzes audiovisual contributions.
- Modeling framework: CAPQ-FAST integrates modality-specific models for faster-playback perceived quality into a unified audiovisual assessment model.The modeling uses TI for video, WPM for speech, and BPM for music as content features.
- Modeling framework: The modeling workflow first compares the developed platform with commercial players, then constructs video and audio quality models and analyzes audiovisual contributions.The stated goal is to derive a content-adaptive model for faster audiovisual playback.
A. Fast Playback Comparison of Media Players
The developed player shows high perceived consistency with mainstream players across video, audio, and audiovisual playback. Chi-square results indicate that playback speed minimally affects perceived consistency under the tested conditions.
- Video playback: Video playback similarity with the proposed player exceeded 88% on average, led by YouTube at 96.39% and followed by Kwai at 93.40%.TikTok, KMPlayer, and VLC reached 90.56%, 90.28%, and 88.06%, respectively.
- Audio playback: Audio playback similarity averaged above 90%, with TikTok, Kwai, and YouTube each exceeding 95% similarity.VLC and KMPlayer were slightly lower, at around 90%.
- Audiovisual playback: Audiovisual playback similarity exceeded 95% with TikTok, Kwai, and YouTube and remained above 90% with VLC and KMPlayer.The proposed player therefore remained highly consistent with all tested mainstream players for audiovisual playback.
- Statistical analysis: Playback speed did not significantly change perceived-consistency distributions for video, audio, or audiovisual content.The reported audiovisual χ² values ranged from 1.605 to 7.162, with p-values between 0.128 and 0.808; the overall conclusion was p > 0.05.
- Overall comparison: Under the experimental conditions, subjective differences between the developed player and target players were small.This supports using the platform for subsequent experiments.
B. Perceived Quality Analysis and Modeling
Perceived quality generally declines as playback speed increases, but the decline depends on modality-specific content characteristics. The paper models these differences using TI, WPM, and BPM as content-adaptive descriptors.
- Video: Video perceived quality decreases with playback speed, with significantly different decline patterns across five video groups.The Friedman test reports χ²=16.00 and p=0.003.
- Video: For high-motion video, quality drops by about 2.49 points at 2x and 3.33 points at 3x playback relative to 1x.Group 5 has Avg TI=27.05; at 3x, ratings fall to roughly 1–2.
- Video: TI captures temporal dynamics and motion complexity, while higher TI is associated with stronger degradation under faster playback.The model treats TI as a descriptor of visual comprehensibility rather than semantic importance, narrative relevance, user interest, or emotional value.
- Speech: Speech quality declines significantly across groups, with slower-speaking speech dropping about 0.54 at 2x and 0.71 at 3x relative to 1x.The Friedman test reports χ²=12.00 and p=0.007; slower speech generally remains in the 4–5 quality range.
- Speech: Faster-speaking speech is less tolerant of acceleration, dropping about 2.9 points at 2x and 3.62 points at 3x relative to 1x.When playback does not exceed 1.5x, ratings are generally 3–5; beyond 2x, perceived quality declines more sharply.
V. PERFORMANCE EVALUATION
The performance evaluation has two aims: validating perceived-quality prediction and examining whether the model helps predict willingness to adopt faster playback.
- Evaluation aims: The evaluation validates the proposed perceived-quality assessment model using prediction performance analysis.The section also examines the model’s contribution to predicting users’ willingness to adopt faster playback in real-world applications.
- Evaluation aims: The evaluation examines prediction performance for perceived quality under faster playback.This is one of the two stated parts of the performance verification.
- Evaluation aims: The evaluation examines the model’s contribution to predicting users’ willingness to adopt faster playback.This is the second stated part of the performance verification.
A. Performance Validation of the Proposed Model
The proposed model is compared with baseline models using subjective ratings and standard correlation and error metrics. It outperforms the baselines, with statistically significant differences in prediction performance.
- Validation setup: Predicted perceived quality is compared with MOS from Session 1 of Experiment 3.The comparison evaluates model predictions against subjective user ratings.
- Baselines: The baselines omit content-specific characteristics of video and audio.Baseline 1 fits one video model across all data, while Baselines 2 and 3 use exponential audio models without content characteristics.
- Metrics: Performance is evaluated using PCC, SROCC, and RMSE, where higher correlation and lower error indicate better performance.These metrics are recommended by VQEG.
- Results: The proposed model outperforms the baselines, with predictions closely aligning with subjective user ratings across video, audio, and audiovisual content.Fig. 7 provides scatter plots comparing MOS with predicted perceived quality.
- Results: All F-test significance values are below 0.05, indicating statistically significant performance differences between the proposed and baseline models.The residual analysis was conducted separately for video, audio, and audiovisual content.
B. Application Validation for Coarse-Grained Playback Intention Interpretation
The validation examines whether perceived quality helps interpret playback intentions across speeds. At higher speeds, users’ EV or SS preferences vary substantially with quality, making perceived quality a useful but coarse-grained auxiliary cue.
- Playback intention validation: At 2x speed, the baseline EV selection rate is 53.0%, while acceptable perceived quality raises average EV selection to 66.1%.When perceived quality falls below 3, EV selection shows a 41% deviation from the baseline, with users favoring SS.
- Playback intention validation: At speeds up to 1.5x, perceived quality remains above 3 and baseline and quality-informed predictions are relatively close.Most users choose EV, averaging 88.8% at 1.25x and 79.8% at 1.5x.
- Playback intention validation: At 3x speed, EV selection averages 44.2% when perceived quality exceeds 3, compared with a 27.1% baseline.The deviation is 17.1% above 3 quality points and increases to 19.7% below 3 points.
- Playback intention validation: Perceived quality provides a useful auxiliary basis for interpreting faster-playback intentions under different quality conditions.The interpretation remains coarse-grained and should be combined with semantic understanding, user profiles, and viewing history.
VI. CONCLUSION
The study links perceived quality under faster playback to content characteristics across video, audio, and audiovisual modalities, and proposes CAPQ-FAST as a content-aware assessment model. Its predicted scores support playback-intention inference and potential adaptive playback and recommendation applications, while current data and feature coverage limit generalizability.
- Conclusion: Perceived quality depends on both playback speed and content characteristics across video, audio, and audiovisual media.Low-motion video and slower speech or music tolerate higher speeds, whereas fast visuals, speech, and rhythm reduce tolerance.
- Conclusion: CAPQ-FAST integrates content-aware observations and temporal features to assess perceived quality across media types and playback speeds.The model uses TI-, WPM-, and BPM-related temporal information in the study’s modality-specific assessment approach.
- Conclusion: Predicted quality scores support inference of playback intentions and provide a basis for adaptive playback control and content recommendation systems.The paper also identifies potential applications in personalized speed recommendations and user-experience evaluation.
- Conclusion: The study’s limited dataset scale, content diversity, and participant diversity constrain model generalizability.The current temporal features also do not capture semantic importance, narrative relevance, or user interest.