TL;DR
This paper tackles the challenge of automatically detecting phonemic aspiration in voiced plosives, a feature common in Indo-Aryan languages but less studied than in unvoiced stops. They propose novel acoustic features, particularly those related to breathy voice quality in the following vowel, and demonstrate their effectiveness in distinguishing aspirated from unaspirated voiced plosives, even generalizing to cross-language pronunciation assessment.
What Problem Does This Paper Address?
Automatic Speech Recognition (ASR) systems often struggle with accurate phone recognition, relying heavily on language models. Improving phone recognition directly, especially for phonetically distinct sounds, could enhance ASR performance, particularly in challenging scenarios like cross-language applications or non-native accent assessment.
The paper focuses on a specific phonetic distinction: phonemic aspiration in voiced plosives. While aspiration in unvoiced plosives is often signaled by release duration (like Voice Onset Time or VOT), this cue is less reliable for voiced aspirated plosives. Instead, the authors hypothesize that the acoustic realization of aspiration in voiced plosives is primarily marked by the breathy voice quality of the following vowel. This breathiness arises from glottal abduction during the plosive closure, which overlaps with the vowel's onset. The challenge is to identify and automatically extract acoustic features that reliably capture this breathiness and other related cues to distinguish between aspirated and unaspirated voiced plosives across different speakers and languages.
Methods
The core of the paper lies in the extraction and evaluation of novel acoustic features designed to capture the subtle acoustic cues of phonemic aspiration in voiced plosives.
- Acoustic Landmark Detection: Before feature extraction, precise temporal landmarks (release burst onset and vowel onset) are crucial. The authors employ a two-stage procedure:
- Coarse Segmentation: Using a broad-class phonetic transcription from an MFCC-HMM ASR system.
- Refinement: The release burst onset is detected by the peak in the rate-of-rise of smoothed energy in a high-frequency band. Vowel onset is detected by a rapid rise in the low-frequency amplitude envelope.
- Feature Extraction: Based on the understanding that breathiness and aspiration noise are key, they extract several features:
- VOT: Standard duration between burst and vowel onset.
- Voice Quality Features: * H1-H2: Ratio of the first two harmonic amplitudes, sensitive to glottal open quotient (OQ) in breathy voice.
- Spectral Tilt: Measures how quickly spectral energy decreases with frequency. They use variations like A1-A3 and a "low band slope" feature.
- B3 Band Energy: Another measure sensitive to spectral tilt.
- Aspiration Noise Features: * SNR: Ratio of harmonic energy to noise energy, estimated using cepstral liftering to isolate source characteristics.
- F1-F3 Sync: Measures the uncorrelatedness of signal components in different frequency bands, capturing aspiration noise.
- Pre-onset Noise Feature (E1-E2): Specifically for affricates, this measures the energy ratio in different frequency bands before vowel onset, capturing the fricated aspiration noise.
- Classification: Gaussian Mixture Models (GMMs) with 6 mixtures and full covariance were trained for a two-way classification task (aspirated vs. unaspirated) on these extracted features.
Datasets
- Marathi Speech Database: Used for training and evaluating the acoustic features and models. It comprises Marathi spoken words with word-initial plosives across various vowel contexts, uttered by 20 native speakers (equal male/female) in carrier sentences.
- Hindi Speech Dataset: Used for pronunciation assessment. This dataset includes:
- 20 native Hindi speakers.
- 10 Tamil L1 speakers (who have had some exposure to Hindi but differ significantly in their native plosive systems).
- Speech was recorded as read-out words containing target phones in word-initial position across vowel contexts, embedded in carrier phrases.
Key Insights and Ideas
- Breathy Voice is Key: The most significant insight is the confirmation that breathy voice quality in the following vowel is a primary acoustic cue for phonemic aspiration in voiced plosives, rather than VOT.
- Multidimensionality of Aspiration: Aspiration is not a single acoustic event but a combination of spectral shape changes (due to breathiness) and aspiration noise. Different features capture different aspects, and combining them leads to better performance.
- Feature Generalization: The acoustic features developed for Marathi demonstrate good generalization to Hindi, suggesting they capture fundamental phonetic properties rather than language-specific idiosyncrasies.
- Automatic vs. Manual Landmarks: While manual landmark detection generally yields better results, the paper highlights that for voiced aspirated affricates, automatic detection can be more reliable due to the inherent difficulty in precise manual annotation of vowel onset.
- Non-native Assessment: The proposed features are more effective than standard MFCCs for distinguishing native from non-native speakers, reinforcing the idea that phonetically motivated features are superior for pronunciation error detection.
Key Results
- Improved Accuracy: Combining spectral shape and noise features significantly improved classification accuracy for distinguishing voiced aspirated from unaspirated plosives. For stops, accuracy reached 85.1% with automatic labels and 86.2% with manual labels using the best feature set (S3). For affricates, accuracy was 77.6% (automatic) and 78.0% (manual) with the added E1-E2 feature.
- Cross-Language Transfer: The trained models achieved good performance on Hindi voiced plosives, with accuracies of 84.9% for stops and 79.2% for affricates, demonstrating cross-language transferability.
- Non-native Speaker Detection: The proposed features showed better accuracy and separation between native and non-native speakers compared to standard MFCC-HMM systems in the pronunciation assessment task.
Limitations and Future Work
- Limitations:
- The paper acknowledges that manual landmark detection is generally better, except for the difficult case of affricate vowel onsets.
- While the features generalize well between Marathi and Hindi, the study is limited to these two languages.
- Future Work:
- The authors plan to extend their work to larger datasets and other Indo-Aryan languages, such as Gujarati, which also exhibits phonemic breathiness in vowels.
- Further investigation into the optimal combination and weighting of features for different phonetic contexts is implied.
Summary updated