TL;DR
This paper investigates how speakers modify their speech in noisy environments (the Lombard effect) to improve intelligibility, focusing on critical announcements like train station alerts. They found that both spectral shaping and prosodic adjustments (like pitch and emphasis) contribute to better understanding in noise, offering a pathway to enhance public address systems.
What Problem Does This Paper Address?
Communicating effectively in noisy environments is a persistent challenge, especially for public announcements where clarity is paramount. While the "Lombard effect" – the involuntary change in speech production when speaking in noise – is known to improve intelligibility, the specific acoustic characteristics responsible, particularly prosodic ones, are not fully understood. This paper aims to quantify these acoustic changes in Marathi train station announcements and apply them to enhance speech intelligibility in noisy listening conditions.
Methods
The researchers analyzed recordings of Marathi train station announcements made under two conditions: normal (quiet) and Lombard (noisy). They used the Fujisaki model to analyze and represent the fundamental frequency (F0) contours, focusing on phrase and accent commands. This allowed them to quantify changes in prosody. For spectral analysis, they calculated spectral shaping filters to transform normal speech into Lombard speech, aiming to replicate the spectral envelope changes. These modified speech signals were then used in listening tests to evaluate intelligibility gains.
Datasets
The study utilized a newly created corpus of 10 Marathi sentences, designed to mimic actual train station announcements. These sentences contained placeholders for train numbers, names, and platform numbers, chosen to create a high cognitive load for listeners. Recordings were made by 3 native Marathi speakers (2 male, 1 female) under both normal and Lombard conditions. For listening tests, synthesized utterances were created using a subset of the recorded data, with added background noise at different Signal-to-Noise Ratios (SNRs).
Key Insights and Ideas
- Beyond Vocal Effort: While increased vocal effort is a hallmark of the Lombard effect, the study highlights that subtle prosodic modifications (intonation, phrasing) are also crucial for intelligibility enhancement.
- Fujisaki Model's Utility: The Fujisaki model proved effective in parameterizing F0 contours, allowing for linguistically valid prosodic modifications that can be applied to speech synthesis.
- Spectral Shaping is Key, but Prosody Helps: Spectral shaping alone significantly improved intelligibility, but combining it with F0 modifications yielded even greater gains, suggesting a synergistic effect.
- Speaker Variability: The magnitude of spectral changes varied considerably between speakers, indicating the need for speaker-dependent modeling for optimal results.
- "Sense of Urgency" in Lombard Speech: The observation that Lombard speech duration marginally decreases, attributed to a "sense of urgency," is an interesting qualitative finding.
Key Results
- Listening tests showed a significant improvement in Word Error Rate (WER) for Lombard-inspired modified speech compared to normal speech in noisy conditions.
- For example, at a -6 dB SNR, the WER for train names dropped from 5.9% to 2.4% with high vocal effort Lombard speech. For train numbers, WER decreased from 25.3% to 13.2% under the same conditions.
- Spectral shaping alone provided substantial intelligibility gains.
- Combining spectral shaping with F0 modifications resulted in the most significant WER improvements, particularly at lower SNRs.
- Prosodic modifications (F0 changes) alone showed a smaller but definite improvement in intelligibility.
Limitations and Future Work
The authors acknowledge that the synthesized F0 modifications sometimes sounded slightly unnatural due to artefacts from the TD-PSOLA algorithm, especially with large F0 scaling. Future work should focus on improving the synthesis of these prosodic modifications. They also suggest replicating these findings on a larger speech and noise dataset and exploring other aspects of clear speech articulation that could further enhance intelligibility. The study also notes that the intelligibility testing was restricted to train numbers for reliability, and further investigation into other utterance components would be beneficial.
Summary updated