TL;DR
This paper investigates how listeners prefer to adjust the "local criterion" (LC) in Ideal Binary Masking (IBM) for speech enhancement, moving beyond just intelligibility. They found significant individual differences in preferred LC values, but a core group of listeners showed consistent preferences that depended on noise type, overall SNR, and sentence difficulty, offering valuable guidance for developing more perceptually optimized speech enhancement systems.
What Problem Does This Paper Address?
Computational auditory scene analysis (CASA) aims to mimic how humans separate sounds in complex environments. A key technique in this field is the Ideal Binary Mask (IBM), which isolates speech by retaining time-frequency regions with a high local signal-to-noise ratio (SNR) and discarding those with low SNR. While previous research has shown that speech intelligibility is generally robust to the specific choice of the local criterion (LC) within a certain range (typically -12 to 0 dB), it's unclear whether listeners have preferences for certain LC values when the intelligibility is already high (at a ceiling). This paper addresses this gap by exploring listeners' subjective preferences for LC settings in IBM-processed speech, aiming to understand what makes the enhanced speech sound "best" rather than just "understandable."
Methods
The core of this study is a real-time implementation of the IBM algorithm. This system allows listeners to interactively adjust the local criterion (LC) using a programmable knob while listening to IBM-processed speech. The process involves:
- Frame-based Processing: Audio is processed in short frames (8 ms).
- Filterbank Analysis: Speech and noise are passed through a 32-channel gammatone filterbank.
- SNR Estimation: The local SNR is estimated for each time-frequency bin.
- IBM Generation: A binary mask is created where a "1" is assigned if the local SNR meets or exceeds the user-adjustable LC, and "0" otherwise.
- Audio Reconstruction: The masked signal is reconstructed and played back.
Listeners were instructed to adjust the LC to achieve the "clearest and easy to listen to for a long time." This self-adjustment method allowed for the measurement of preferred LC values under various conditions. A hierarchical cluster analysis was used to group listeners based on their preferred LC patterns across different experimental conditions.
Datasets
- Speech Corpora:
- 250 sentences from the Hearing in Noise Test (HINT) corpus.
- First 250 sentences from the IEEE corpus.
- Both corpora were concatenated into long audio files for continuous playback.
- Background Noises:
- Airplane noise and train noise from the ESC-50 database.
- 10-talker babble from the AzBio database.
No new datasets were introduced; existing, well-established corpora and noise databases were used.
Key Insights and Ideas
- Beyond Intelligibility: The study highlights that while IBM processing might achieve near-perfect intelligibility across a range of LCs, listeners have distinct preferences that go beyond just understanding the words. This suggests that factors like listening effort, perceived quality, and distortion play a role in subjective preference.
- Significant Individual Differences: There's a wide variation in preferred LC values among listeners, indicating that a one-size-fits-all approach to setting LC might not be optimal for perceptual quality.
- Clustering Reveals Patterns: A cluster analysis successfully grouped listeners, with a substantial group (11 out of 20) showing consistent patterns in their preferences. This suggests that while individual differences exist, there are also common perceptual strategies.
- LC Sensitivity to Context: For the "typical" listener group, preferred LC values are not arbitrary but are influenced by:
- Noise Type: Different background noises lead to different preferred LCs.
- Overall SNR: As the overall SNR increases, listeners tend to prefer higher LCs.
- Speech Material Difficulty: More difficult speech materials (like IEEE sentences compared to HINT) lead to higher preferred LCs. This implies listeners prioritize preserving more speech information when the speech itself is harder to process.
- No Preference for High LCs: No listener preferred an LC value above 0 dB, reinforcing the idea that excessively high LCs, which remove too much speech information, are perceptually undesirable.
Key Results
- Large Individual Differences: Preferred LC values varied widely, with some listeners preferring very low LCs (below -40 dB) and others preferring values closer to 0 dB.
- Six Listener Groups Identified: A hierarchical cluster analysis identified six distinct groups of listeners based on their preferred LC patterns.
- Group 1 Dominance: Group 1, comprising 11 out of 20 listeners, showed consistent patterns. For this group:
- Preferred LC values were generally higher for IEEE sentences than HINT sentences.
- Preferred LC values increased with increasing overall SNR (from -5 dB to 5 dB).
- A statistically significant main effect of speech material and overall SNR was found for Group 1.
- The effect of noise type was marginally significant.
- Discrepancies in Other Groups: Some smaller listener groups showed large discrepancies in preferred LCs between the two speech materials, suggesting potential unreliability in their estimates.
Limitations and Future Work
- Limitations:
- The study acknowledges large individual differences, meaning that the findings from the dominant "Group 1" might not generalize to all listeners.
- The "typical" listener group's preference for higher LCs with more difficult speech materials suggests a trade-off: listeners might accept more noise to preserve speech cues when the speech is inherently harder to understand. This could imply increased listening effort, which wasn't directly measured.
- While the study found that preferred LC scales with overall SNR, the effect of noise type was less clear for airplane noise compared to babble and train noise, suggesting that other noise characteristics might also influence preferences.
- Future Work:
- The authors suggest that these listener-preferred LC values could be more appropriate targets for speech enhancement algorithms than arbitrarily chosen values.
- Future research could explore the relationship between preferred LC values and objective measures of listening effort or subjective speech quality more directly.
- Investigating the influence of other noise characteristics beyond SNR and type on listener preferences would be valuable.
- Further analysis of the smaller listener groups might reveal distinct perceptual strategies or issues with the experimental setup for those individuals.
Summary updated