TL;DR
This paper investigates whether automatically detecting a user's chronological age is necessary for spoken dialogue systems (SDS) to adapt to user needs. The authors demonstrate that while acoustic features alone are poor predictors of age-related needs, simple lexical features significantly improve age prediction and are even more effective at directly predicting crucial user characteristics like information processing speed and interaction style. The core takeaway is that SDS should prioritize adapting to observed user behavior rather than relying on potentially unreliable age estimations.
What Problem Does This Paper Address?
Spoken dialogue systems (SDS) often aim to adapt their behavior to better serve users. A common approach is to infer user characteristics like age, gender, or emotion from speech to tailor the system's responses. This paper focuses on the specific challenge of adapting SDS for older users, who may experience age-related changes in cognition, speech, and interaction preferences. The authors question the prevailing assumption that accurately predicting a user's chronological age is the best or even a necessary first step for enabling such adaptations. They highlight the inherent diversity within the "older user" demographic, making a single age-based adaptation strategy problematic. The central problem is to determine if direct prediction of relevant user characteristics is more effective than inferring them indirectly through age recognition.
Methods
The researchers employed a combination of acoustic and lexical features to predict user characteristics.
- Acoustic Features: Standard features like Mel Frequency Cepstral Coefficients (MFCCs), pitch-related measures (F0, jitter, shimmer), vocal tract length normalization (VTLN) factor, and speaking rate were extracted from speech utterances.
- Lexical Features: These were derived from word groups identified by keyword spotters, including categories like positive/negative feedback, greetings, apologies, and hesitations. The frequency of these word groups was used as a predictor.
- Regression Models: Linear regression was used for the continuous target variable (information processing speed, DSST), and logistic regression for binary variables (interaction style, IS, and age group).
- Feature Selection: To combat overfitting with a small dataset, stepwise feature selection (using AIC) was employed to identify the most predictive features for each model (Voice, Lex, and Voice+Lex).
- Evaluation: Leave-one-out cross-validation was used for acoustic features, while for MFCCs, a speaker-independent approach was adopted.
Datasets
The study utilized a dataset of 448 appointment scheduling dialogues collected using a Wizard-of-Oz paradigm.
- Participants: 50 participants were involved, divided into two groups: 26 older adults (average age 66) and 24 younger adults (average age 22).
- Annotation: Participants underwent cognitive tests, and their interaction styles were analyzed. The dataset is a subset of previously collected data described in references [4] and [16].
Key Insights and Ideas
- Age is a Poor Proxy: Chronological age, as estimated from vocal cues, is not a reliable indicator of the specific needs of older users for adaptation in SDS.
- Lexical Features Shine: Simple lexical features (word groups) are surprisingly powerful predictors of crucial user characteristics, often outperforming acoustic features.
- Direct Prediction is Superior: Models that directly predict information processing speed (DSST) and interaction style (IS) perform significantly better than models that rely on predicted age.
- Behavior Over Demographics: The findings strongly suggest that SDS should adapt based on observed user behavior and characteristics (like interaction style or cognitive load indicators) rather than solely on estimated chronological age.
- Combining Features Helps: While lexical features are potent on their own, combining them with acoustic features further enhances performance, particularly for age prediction.
Key Results
- Age Prediction: The best age prediction model (Voice+Lex) achieved 90% accuracy, 92.31% precision, and 88.89% recall, significantly outperforming acoustic-only (Voice: 70% Acc, 84.62% Prec, 66.67% Recall) and MFCC-based models. Lexical features alone also showed strong performance (82% Acc, 76.92% Prec, 86.96% Recall).
- Interaction Style (IS) Prediction: A dedicated IS model using lexical features achieved near-perfect performance (100% Accuracy, Precision, and Recall). Age-based models failed to improve on the baseline.
- Information Processing Speed (DSST) Prediction: The baseline RMS error was 15.50. Using predicted age reduced this to 10.23. The best model (Voice+Lex) achieved an RMS error of 9.77, which was only slightly better than the model using predicted age (11.34), but significantly better than voice-only (12.29) or lex-only (12.20) models.
Limitations and Future Work
- Data Size: The study acknowledges that the dataset of 50 participants is relatively small, making it vulnerable to overfitting, which is why feature selection was crucial.
- Time-to-Adaptation: A limitation is that some lexical features (e.g., hesitations) can only be reliably detected after the dialogue has begun, whereas some systems aim to adapt very early. Future work could involve compiling lexical statistics over time within the dialogue history.
- Task Specificity: While the lexical features used are not strictly task-specific, further investigation is needed to confirm their transferability to other tasks and SDS.
- Feasibility of Direct Prediction: The authors note that while direct prediction is often preferable, it may not always be feasible or desirable (e.g., for age-specific product recommendations).
Summary updated