Problem: Phonetic transcription is crucial in diagnosing speech sound disorders (SSDs) but is susceptible to transcriber experience and perceptual bias. Current forced alignment (FA) tools, which annotate audio files to determine spoken content and its placement, often require manual transcription, limiting their effectiveness. Method: We introduce a novel, text-independent forced alignment model that autonomously recognises individual phonemes and their boundaries, addressing these limitations. Our approach leverages an advanced, pre-trained wav2vec 2.0 model to segment speech into tokens and recognise them automatically. To accurately identify phoneme boundaries, we utilise an unsupervised segmentation tool, UnsupSeg. Labelling of segments employs nearest-neighbour classification with wav2vec 2.0 labels, before connectionist temporal classification (CTC) collapse, determining class labels based on maximum overlap. Additional post-processing, including overfitting cleaning and voice activity detection, is implemented to enhance segmentation. Results: We benchmarked our model against existing methods using the TIMIT dataset for normal speakers and, for the first time, evaluated its performance on the TORGO dataset containing SSD speakers. Our model demonstrated competitive performance, achieving a harmonic mean score of 76.88% on TIMIT and 70.31% on TORGO. Implications: This research presents a significant advancement in the assessment and diagnosis of SSDs, offering a more objective and less biased approach than traditional methods. Our model\'s effectiveness, particularly with SSD speakers, opens new avenues for research and clinical application in speech pathology.

Impressionistic coding of sociolinguistic variables like English (ING), the alternation between pronunciations like talkin\' and talking, has been a central part of the analytic workflow in studies of language variation and change for over a half-century. Techniques for automating the measurement and coding for a wide range of sociolinguistic data have been on the rise over recent decades but procedures for coding some features, especially those without clearly defined acoustic correlates like (ING), have lagged behind others, such as vowels and sibilants. This paper explores computational methods for automatically coding variable (ING) in speech recordings, examining the use of automatic speech recognition procedures related to forced alignment (using the Montreal Forced Aligner) as well as supervised machine learning algorithms (linear and radial support vector machines, and random forests). Considering the automated coding of pronunciation variables like (ING) raises broader questions for sociolinguistic methods, such as how much different human analysts agree in their impressionistic codes for such variables and what data might act as the \"gold standard\" for training and testing of automated procedures. This paper explores several of these considerations in automated, and manual, coding of sociolinguistic variables and provides baseline performance data for automated and manual coding methods. We consider multiple ways of assessing algorithms\' performance, including agreement with human coders, as well as the impact on the outcome of an analysis of (ING) that includes linguistic and social factors. Our results show promise for automated coding methods but also highlight that variability in results should be expected even with careful human coded data. All data for our study come from the public Corpus of Regional African American Language and code and derivative datasets (including our hand-coded data) are available with the paper.

This paper discusses how the transcription hurdle in dialect corpus building can be cleared. While corpus analysis has strongly gained in popularity in linguistic research, dialect corpora are still relatively scarce. This scarcity can be attributed to several factors, one of which is the challenging nature of transcribing dialects, given a lack of both orthographic norms for many dialects and speech technological tools trained on dialect data. This paper addresses the questions (i) how dialects can be transcribed efficiently and (ii) whether speech technological tools can lighten the transcription work. These questions are tackled using the Southern Dutch dialects (SDDs) as case study, for which the usefulness of automatic speech recognition (ASR), respeaking, and forced alignment is considered. Tests with these tools indicate that dialects still constitute a major speech technological challenge. In the case of the SDDs, the decision was made to use speech technology only for the word-level segmentation of the audio files, as the transcription itself could not be sped up by ASR tools. The discussion does however indicate that the usefulness of ASR and other related tools for a dialect corpus project is strongly determined by the sound quality of the dialect recordings, the availability of statistical dialect-specific models, the degree of linguistic differentiation between the dialects and the standard language, and the goals the transcripts have to serve.

Vowel durations are most often utilized in studies addressing specific issues in phonetics. Thus far this has been hampered by a reliance on subjective, labor-intensive manual annotation. Our goal is to build an algorithm for automatic accurate measurement of vowel duration, where the input to the algorithm is a speech segment contains one vowel preceded and followed by consonants (CVC). Our algorithm is based on a deep neural network trained at the frame level on manually annotated data from a phonetic study. Specifically, we try two deep-network architectures: convolutional neural network (CNN), and deep belief network (DBN), and compare their accuracy to an HMM-based forced aligner. Results suggest that CNN is better than DBN, and both CNN and HMM-based forced aligner are comparable in their results, but neither of them yielded the same predictions as models fit to manually annotated data.