Abstract
Unsupervised word segmentation aims to identify word boundaries in continuous speech without the aid of manual transcriptions. This paper introduces GradSeg, a method that leverages the temporal dynamics of self-supervised audio embeddings. By observing that feature changes are minimal within words, GradSeg generates pseudo-labels from "gradient valleys" to train a simple linear classifier. Despite its simplicity, GradSeg significantly outperforms existing probabilistic models on standard benchmarks like the Buckeye and YOHO corpora, demonstrating remarkable data efficiency and achieving a new state-of-the-art in the field.
Introduction
Spoken language lacks the explicit delimiters found in written text, presenting a significant challenge for unsupervised word segmentation. This task is essential for modeling language acquisition and developing tools for low-resource languages where labeled datasets are unavailable. Traditional approaches often rely on complex probabilistic frameworks or multi-stage pipelines that first identify phonemes before grouping them into words. GradSeg simplifies this process by exploiting the geometric properties of self-supervised feature spaces, moving away from complex probability toward a simple observation about temporal gradients.
Methods
GradSeg operates on a straightforward pipeline designed to extract boundary information from raw audio:
- Feature Extraction: Raw audio is processed through a pre-trained, frozen Wav2Vec 2.0 model to obtain high-dimensional embeddings every 20ms.
- Temporal Gradient Calculation: The Euclidean distance between subsequent embeddings is computed. Low gradient magnitudes indicate stable regions, while peaks indicate rapid change.
- Pseudo-Labeling: The core innovation lies in identifying "valleys"—frames with the lowest 20% of gradient values. These are used as pseudo-labels to mark frames that are "definitely not a boundary."
- Linear Classification: A simple Ridge Regression model is trained to map the audio features to these pseudo-labels.
- Boundary Prediction: At inference, the classifier provides a "wordness" score for every frame. Final boundaries are determined using Non-Maxima Suppression (NMS) based on a rough estimate of the average word duration in the target language.
Key Insights and Data Efficiency
A critical finding of this research is that "valleys" in the gradient signal are more reliable indicators of word interiors than "peaks" are of boundaries. Furthermore, the paper confirms that self-supervised models like Wav2Vec 2.0 already capture boundary information in their latent space; GradSeg simply provides an unsupervised mechanism to extract it.
GradSeg is also extremely data-efficient. The model reaches peak performance with as few as 100 training utterances. Adding more data beyond this point does not significantly improve the F1-score, making the method highly suitable for low-resource environments.
Results
GradSeg achieved a new state-of-the-art in unsupervised word segmentation across two standard benchmarks:
- Buckeye Corpus: It achieved an F1-score of 44.1%, a massive jump over the previous best of 32.1% (Buckeye_SCPC).
- YOHO Dataset: It achieved an F1-score of 43.8%, outperforming the previous best (DSegKNN) at 42.9%.
Beyond accuracy, GradSeg is significantly more efficient than previous probabilistic or nearest-neighbor-based approaches due to its use of a frozen feature extractor and a simple linear classifier.
Limitations and Future Work
While powerful, GradSeg depends on high-quality self-supervised models pre-trained on large amounts of unlabeled audio. Its performance on extremely low-resource or "exotic" languages may depend on the generalization capabilities of the underlying pre-trained model. Additionally, the method is sensitive to the NMS hyperparameters related to word duration. Future research may explore more sophisticated peak-finding algorithms and alternative classification objectives to further refine boundary detection.
Summary updated