TL;DR
This study investigates how different speech styles, beyond just "clear speech," affect intelligibility in noise. They found that increased speech duration, particularly in computer-directed speech, significantly boosts intelligibility, while changes in fundamental frequency (F0) have a more complex and less predictable role.
What Problem Does This Paper Address?
When people speak to those with hearing impairments, in noisy environments, or to non-native speakers, they often adopt a "clear speech" style. This style is known to improve intelligibility, and it's characterized by prosodic changes like slower speech rate and altered fundamental frequency (F0). However, the precise mechanisms by which these prosodic modifications contribute to intelligibility are not fully understood. Specifically, the paper aims to disentangle the effects of different prosodic changes (like duration, pauses, and F0 variations) across various speech styles, moving beyond the general "clear speech" concept to see which specific acoustic features are most beneficial for listeners in noisy conditions.
Methods
The researchers recorded a native British English speaker producing 25 TIMIT sentences in five distinct speech styles:
- Plain speech: Baseline conversational style.
- Infant-directed speech: Elicited by instructions to speak to an infant.
- Computer-directed speech: Elicited by instructions to speak to a computer.
- Foreigner-directed speech: Elicited by instructions to speak to a non-native speaker.
- Shouted speech: Produced while listening to loud babble noise.
These recorded sentences were then mixed with 9-speaker babble noise at a 0 dB Signal-to-Noise Ratio (SNR).
Two groups of listeners evaluated the intelligibility:
- Laboratory-based listeners: 20 native English speakers in controlled listening booths.
- Amazon Mechanical Turk (AMT) listeners: 60 self-reported native English speakers, instructed to listen in a quiet environment.
Acoustic analysis involved:
- Manual segmentation: Sentences were segmented at the phoneme level.
- Prosodic parameter extraction: Measurements of sentence duration, segment duration, and unfilled pause duration were taken.
- F0 analysis: Mean F0 and F0 range were calculated for voiced segments.
- Glimpsing analysis: A model was used to estimate the proportion and count of time-frequency regions where speech was more energetic than the masker, simulating how listeners might extract information in noise.
Datasets
The study used the TIMIT dataset for sentence selection. The specific sentences were produced by a single male native British English speaker in five different styles. The noise used was a 9-speaker babble-shaped noise from a previously published source [20]. No new benchmarks were introduced; the focus was on analyzing existing speech in noise.
Key Insights and Ideas
- Duration is King (for intelligibility): The most significant finding is that increased speech duration, particularly evident in computer-directed speech, strongly correlates with improved intelligibility. This suggests that listeners benefit from more time to process the speech signal when it's stretched out.
- F0's Role is Nuanced: While F0 median and range varied significantly across styles, their direct impact on intelligibility was less clear-cut. Infant-directed speech, which had a higher F0 range, was less intelligible, suggesting that simply increasing F0 range isn't always beneficial. Conversely, computer-directed speech, with a flatter F0 contour, was highly intelligible. This challenges simpler models of F0's contribution.
- Infant-Directed Speech is a Curveball: Surprisingly, infant-directed speech, often considered a form of "clear speech," was less intelligible than plain speech. This highlights that not all modifications associated with child-directed speech automatically enhance intelligibility in noisy adult listening scenarios.
- Glimpse Count Matters More Than Proportion: The "glimpsing" analysis revealed that the total number of time-frequency regions where speech energy exceeded noise (glimpse count) was a good predictor of word identification, rather than the proportion of such regions. This suggests that having more opportunities to "glimpse" the signal, even if the overall proportion of clear signal isn't drastically higher, is key.
- Segment Elongation Drives Glimpse Count: The study posits that segment elongation, a key feature of computer-directed speech, directly leads to an increased glimpse count, thereby boosting intelligibility.
Key Results
- Computer-directed speech was the most intelligible across both listener groups.
- Infant-directed speech was the least intelligible.
- Sentence duration was significantly longer in all non-plain speech styles except infant-directed speech. Computer-directed sentences were, on average, 68.4% longer than plain speech sentences.
- There was a strong negative correlation between sentence duration and Word Error Rate (WER) (ρ = -0.95, p < .05).
- F0 median was significantly higher in all non-plain speech styles compared to plain speech, except for computer-directed speech.
- F0 range was significantly higher in infant-directed speech but significantly lower in computer-directed speech compared to plain speech.
- The glimpse count correlated well with word identification rate (ρ = 0.94, p < .05).
Limitations and Future Work
- Single Talker: The study used only one male native British English speaker. This limits the generalizability of the findings to other talkers, genders, or accents.
- Acted Styles: The speech styles (except shouted) were elicited via instructions rather than naturally occurring in specific communicative contexts. This means the acoustic characteristics might not perfectly reflect real-world "infant-directed" or "computer-directed" speech.
- F0 Complexity: The paper acknowledges that the role of F0 in intelligibility is more complex than previously thought and warrants further investigation.
- Other Modifications: The authors mention that other articulatory-acoustic changes (e.g., consonant hyper-articulation, vowel space expansion) might have been present in the acted styles and could have influenced intelligibility, but these were not the primary focus.
- Future Directions: The authors suggest further research is needed to fully describe the relationship between prosodic adjustments and intelligibility, particularly exploring the nuanced role of F0 and the impact of other articulatory changes. They also imply that understanding how listeners attend to these prosodic cues is an open question.
Summary updated