Comparison of Non-Parametric Bayesian Mixture Models for Syllable Clustering and Zero-Resource Speech Processing

Abstract

Zero-resource speech processing (ZS) systems aim to learn structural representations of speech without access to labeled data. A starting point for these systems is the extraction of syllable tokens utilizing the rhythmic structure of a speech signal. Several recent ZS systems have therefore focused on clustering such syllable tokens into linguistically meaningful units. These systems have so far used heuristically set number of clusters, which can, however, be highly dataset dependent and cannot be optimized in actual unsupervised settings. This paper focuses on improving the flexibility of ZS systems using Bayesian non-parametric (BNP) mixture models that are capable of simultaneously learning the cluster models as well as their number based on the properties of the dataset. We also compare different model design choices, namely priors over the weights and the cluster component models, as the impact of these choices is rarely reported in the previous studies. Experiments are conducted using conversational speech from several languages. The models are first evaluated in a separate syllable clustering task and then as a part of a full ZS system in order to examine the potential of BNP methods and illuminate the relative importance of different model design choices.

Speech Sequence Embeddings using Nearest Neighbors Contrastive Learning

“…Figure 2: NED/COV curves on Zerospeech corpora. Our method is represented with a line of blue dots and each other competitors as grey points: Garcia-Granada (A [33]), Jansen (B, [6]), Räsänen (L,M,N [7] and D [8] and G,H [9]) , Kamper (O [4] and P [34]), Lynsinski (I,L,J [10]), Bhati (E,F [11]).…”

“…However, such representations are still not good enough for other downstream tasks that need to handle larger word-like units that typically span larger time windows (100ms-1000ms) and have varying duration depending on speech rate. These tasks include word segmentation [4,5] term discovery [6,7,8,9,10,11] and unsupervised alignment of text and speech [12,13] Despite recent research efforts into developing word embeddings also known as speech sequence embeddings (SSE), [14,15,16,17,18] their performances are low and difficult to compare for lack of a common evaluation benchmark.…”