Acoustic characterization of children with speech delay

H. Timothy Bunnell, James B. PolikoffInterspeech 2006

TL;DR

This paper explores using Hidden Markov Models (HMMs), a technique common in speech recognition, to analyze speech patterns in children with developmental speech delays. By adapting normal-speech trained HMMs, they aim to identify acoustic distortions and characterize differences among these children, offering a more efficient and scalable approach than traditional manual phonetic analysis.

What Problem Does This Paper Address?

The field of speech disorders research has lagged in adopting advanced speech analysis and pattern recognition techniques, despite their widespread use in general speech recognition. Current "best practice" in acoustic phonetic analysis relies heavily on labor-intensive, manual methods like phonetic transcription and alignment, which limit sample sizes and scalability. This manual approach also struggles to account for the temporal nature of speech production and can miss subtle acoustic distinctions due to small sample sizes. The authors aim to bridge this gap by applying HMMs, which inherently capture temporal information, to analyze speech from children with developmental delays, enabling more efficient and comprehensive characterization.

Methods

The core of this paper lies in adapting Hidden Markov Models (HMMs) for analyzing speech from children with developmental delays.

  • Normative HMMs: The researchers first trained HMMs on a large corpus of speech from typically developing children (ages 6-8). These "normative HMMs" were designed to capture the acoustic properties of phonetic segments, including their temporal evolution. This approach moves beyond traditional methods that rely on identifying specific acoustic landmarks within a segment.
  • Classification of Disordered /r/ Productions: To characterize speech delays, they focused on the production of the /r/ sound. They used a k-means clustering algorithm to group individual /r/ tokens based on HMM parameters (log likelihood, frame counts, state-wise log likelihood). Subsequently, hierarchical clustering was employed to group talkers based on the distribution of their /r/ tokens across these identified clusters. This allowed them to characterize individual children's speech patterns.

Datasets

  • Normative Data: A dataset of 18,566 words from 208 typically developing children (ages 6-8) was used to train the normative HMMs. This corpus was collected using a portable recording apparatus and included a wide range of vocabulary, some of which was considered advanced for the age group.
  • Disordered Speech Data: Speech stimuli for the disordered group consisted of 1,909 single-word utterances. This set included words from a speech database of normally speaking children and a smaller set of probe words used in a speech training evaluation study for children with syllable-initial /r/ delays.

Key Insights and Ideas

  • HMMs as a Scalable Tool: The paper demonstrates that HMMs, a standard in speech recognition, can be effectively adapted for clinical speech analysis, offering a more efficient alternative to manual transcription and measurement.
  • Temporal Capture: HMMs inherently model the temporal evolution of phonetic segments, obviating the need to pre-define specific measurement locations within a sound, which is a significant advantage over traditional methods.
  • Acoustic Phenotypes: The clustering of talkers based on their HMM-derived /r/ production patterns suggests that these acoustically-based characterizations can serve as "acoustic phenotypes" for speech-delayed children, potentially aiding in understanding the underlying mechanisms of their delays.
  • Subtle Distinctions: The token clustering revealed distinct categories of /r/ acoustic structure, highlighting that even within a group of children with speech delays, there can be varied articulatory strategies.

Key Results

  • Token Clustering: Three distinct clusters of /r/ tokens were identified, showing significant differences in the distribution of tokens from normal vs. disordered speakers. Cluster 1 was predominantly normal (66.2%), Cluster 2 had a mixed distribution, and Cluster 3 was predominantly disordered (78.7%).
  • Talker Clustering: Hierarchical clustering of individual talkers revealed groupings and subgroupings of disordered children, suggesting that their speech patterns could be characterized and differentiated. The composite normal speaker did not directly pair with any individual disordered talker, linking at a moderate dissimilarity level.
  • Efficiency: The HMM-based approach requires minimal "hands-on" data manipulation compared to traditional methods.

Limitations and Future Work

  • Model Structure: The authors acknowledge that further exploration is needed regarding alternative HMM structures, such as varying the number of states or transition rules.
  • Continuous HMMs: The study used discrete HMMs, and the authors suggest that continuous HMMs would likely better capture the variability in both normal and disordered /r/ productions.
  • Snapshot vs. Trajectory: The data represents an "average picture" of each child's performance over a period of improvement, rather than a fixed snapshot of their articulatory strategies.
  • Future Directions: The authors are continuing this work as part of a larger study to explore using acoustic and other factors to characterize speech-delayed children for genetic linkage studies.

Summary updated

BibTeX
@inproceedings{bunnell2006acoustic,
  author = {H. Timothy Bunnell and James B. Polikoff},
  title = {{Acoustic characterization of children with speech delay}},
  booktitle = {Interspeech},
  year = {2006},
}
Figure