TL;DR
This paper investigates how age affects our ability to understand speech over the telephone, comparing natural and synthetic voices. They found that comprehension degrades with age, and synthetic speech is generally harder to understand than natural speech, highlighting a critical challenge for designing accessible communication systems for older adults.
What Problem Does This Paper Address?
As our population ages, there's a growing need for accessible communication technologies, particularly for older adults who may rely more heavily on telephones for information access. While it's widely believed that hearing and understanding spoken language degrades with age, much of the evidence is anecdotal. This paper aims to provide empirical evidence for this phenomenon, specifically focusing on speech comprehension over the telephone, and to explore how this degradation differs between natural and synthetic speech. The motivation is to inform the development of more effective telephone-based information presentation systems and dialogue systems for the elderly.
Methods
The researchers designed a straightforward experiment to measure speech comprehension. Participants listened to utterances over the telephone and were asked to write down what they heard. They used four different voice conditions:
- NN (Natural Speech): A standard natural spoken utterance by a female speaker.
- NS (Natural Speech - Slowed/Articulated): The same female speaker, but instructed to speak slower and more articulately, as if the listener had trouble understanding.
- SN (Synthetic Speech - Predicted Prosody): Synthetic speech generated using a standard diphone synthesis voice with predicted prosody.
- SS (Synthetic Speech - Natural Prosody): Synthetic speech using the same diphone voice, but with prosody (timing, intonation, phrasing) extracted from the NS condition.
Two types of utterances were used: "time" (more constrained and predictable) and "words" (less constrained). The experiment was designed to be short and simple to accommodate elderly participants.
Datasets
Not explicitly discussed in the paper as distinct datasets. The study used utterances recorded from a graduate student (for natural speech) and generated synthetic speech using the Festival Speech Synthesis System with the us1 mbrola voice. The "tokens" used for evaluation were pairs of words or times, with specific examples like "rose bush," "Rose Bowl," "holiday season," and "holiday shopping."
Key Insights and Ideas
- Age-Related Degradation is Real: The study provides quantitative evidence that speech comprehension over the telephone decreases significantly with age.
- Synthetic Speech is a Bigger Hurdle: Synthetic speech, even with natural prosody, is consistently harder to understand than natural speech, especially for older listeners.
- Predictability Matters: Utterances with more constrained semantic and lexical possibilities (like "time") are easier to understand than those with more open-ended possibilities (like "words"), and this holds true across age groups.
- Natural Prosody Helps: The SS condition, which incorporated natural prosody, showed better results than the SN condition, suggesting that natural timing and intonation are crucial for understandability, even in synthetic speech.
- "Trying Harder" Isn't Always Better: The NS condition (slowed/articulated natural speech) did not perform significantly better than the standard NN condition, indicating that simply increasing volume or articulation might not be the most effective strategy for improving comprehension for older adults.
Key Results
- Overall comprehension decreased from 100% for the 20-29 age group to just under 60% for the 80-89 age group.
- Synthetic speech was generally harder to understand than natural speech across all age groups.
- "Time" utterances were understood more often than "words" utterances within the same condition.
- The correlation between age and percentage of correct responses was 0.33, indicating that older individuals generally understood less.
Limitations and Future Work
- Sample Size: The authors acknowledge that the study's findings are preliminary due to a small sample size (67 subjects). More data is needed for definitive conclusions.
- Subject Demographics: The participants were primarily Carnegie Mellon alumni attending Homecoming, a group that is likely more educated and mobile than the general elderly population. Future work should include participants from activity centers and assisted living communities.
- Synthetic Voice Quality: While they used a standard diphone voice, they note that more recent synthesis techniques might yield different results.
- Future Directions: The authors suggest further research into specific prosodic elements (timing, intonation, phrasing) that can improve synthetic speech understandability. They also emphasize the need to improve natural synthesis quality to better serve the aging population.
Summary updated