Abstract: A vast majority of the world's 7,000 spoken languages are predicted to become extinct within this century, including the endangered language of Ladin from the Italian Alps. Linguists who work to preserve a language's phonetic and phonological structure can spend hours transcribing each minute of speech from native speakers. To address this problem in the context of Ladin, our paper presents the first analysis of speech representations and machine learning models for classifying 32 phonemes of Ladin. We experimented with a novel dataset of the Fascian dialect of Ladin, collected from native speakers in Italy. We created frame-level and segment-level speech feature extraction approaches and conducted extensive experiments with 8 different classifiers trained on 9 different speech representations. Our speech representations ranged from traditional features (MFCC, LPC) to features learned with deep neural network models (autoencoders, LSTM autoencoders, and WaveNet). Our highest-performing classifier, trained on MFCC representations of speech signals, achieved an 86% average accuracy across all Ladin phonemes. We also obtained average accuracies above 77% for all Ladin phoneme subgroups examined. Our findings contribute insights for learning discriminative Ladin phoneme representations and demonstrate the potential for leveraging machine learning and speech signal processing to preserve Ladin and other endangered languages.
The Challenge of Language Preservation
Linguists estimate that up to 90% of the world’s 7,000 languages may go extinct by the end of the century. Preserving these languages requires documenting their phonetic structure, but manual transcription is an agonizingly slow process: one minute of recorded speech can take a trained linguist up to two hours to transcribe.
While modern AI has mastered speech recognition for "high-resource" languages like English, these models typically require thousands of hours of data. Endangered languages like Ladin are "low-resource," meaning there is very little labeled data available to train models. This research investigates whether machine learning can still be effective in these data-scarce environments to help linguists break the transcription bottleneck.
Methodology
The study explored two main variables: speech representations (how the computer "hears" the audio) and classification models (how the computer labels the sounds).
1. Speech Representations
Researchers compared traditional features like Mel-frequency cepstral coefficients (MFCC) and Linear Predictive Coding (LPC) against features learned by deep unsupervised models, including standard Autoencoders, LSTM Autoencoders, and a pre-trained WaveNet.
2. Feature Extraction Levels
Two approaches were tested to capture acoustic data:
- Frame-level: Extracting features from standard 10ms windows.
- Segment-level: A more granular approach using 2.5ms sliding windows averaged across a frame to capture finer acoustic nuances.
3. Classifiers
Eight different models were evaluated, ranging from classical algorithms like Support Vector Machines (SVM), Random Forest (RF), and Logistic Regression (LR) to a 2-layer Dense Neural Network (DNN).
Datasets
The study introduces a novel dataset of the Fascian dialect of Ladin, collected during field research in Italy. The total audio consists of 5 hours and 38 minutes of conversations from four native speakers. A subset of roughly 8 minutes (2,520 segments) was meticulously transcribed by linguists to serve as the "ground truth" for training and testing across 33 total categories (32 phonemes plus one for silence).
Key Results and Insights
The experimental results revealed several critical findings for low-resource language processing:
- Traditional Features Still Reign in Data Scarcity: Interestingly, the best-performing model used traditional MFCC representations rather than deep-learned features. This suggests that when training data is extremely limited, well-engineered traditional features are more robust than features learned from scratch by complex neural networks.
- The Power of "Segment-Level" Analysis: The "segment-level" approach consistently outperformed the standard frame-level approach. This indicates that for phoneme classification, capturing the rapid, subtle transitions within a single sound is crucial.
- Top Performance: The combination of Segment-level MFCCs and a Dense Neural Network achieved the highest average accuracy of 86% across all 32 phonemes.
- Acoustic Physics Impacts Accuracy: The model was significantly better at identifying unrounded vowels (90% accuracy) than rounded vowels (80%). Lip rounding lengthens the vocal tract and lowers frequencies, making the resulting spectral envelope less discriminative for the model.
Limitations and Future Work
The primary limitation is the small sample size, with data from only four speakers and less than ten minutes of transcribed audio. This makes the model sensitive to individual speaker characteristics.
Future work will focus on Data Augmentation to expand the training set and Cross-Lingual Transfer Learning. Since Ladin shares phonetic traits with non-endangered languages like Italian, "pre-training" a model on Italian and then "fine-tuning" it on Ladin could significantly boost performance. The ultimate goal is to integrate these models into field tools that generate "draft" transcriptions for linguists to verify, drastically accelerating preservation efforts.
Summary updated