Abstract

Automatic syllable stress detection traditionally relies on heuristic features like energy, duration, and pitch. Recent self-supervised models, such as wav2vec 2.0, capture stress cues more effectively, enhancing scalability. These models encode different information across layers, yet existing research focuses mainly on final-layer features, leaving intermediate layers unexplored. Stress detection is particularly challenging for non-native speakers due to deviations from correct stress patterns. This study analyzes all layers of self-supervised models to identify the most effective for stress detection. We evaluate wav2vec 2.0 (base, large, xlsr) on supervised and unsupervised classifiers using the ISLE corpus of non-native German and Italian speakers. Results show that wav2vec 2.0-large achieves the highest accuracy (96.43% for German, 96.68% for Italian) at layers 16 and 13, respectively, improving by 2.9% and 3.3% over the baseline.

Abstract

Grammar error detection in one's speech, referred to as spoken grammar error detection (SGED), is a critical component for building computer-assisted language learning systems. Traditionally, ASR followed by text-based GED has been used, but ASR errors can limit SGED performance. In this work, we analyse SGED under three speaking conditions: spontaneous, semi-spontaneous, and memorized, using five state-of-the-art ASRs. We examine the decoded text and ASR cues such as confidence scores and aligner probabilities, out of which, for the latter one, the ground-truth grammatically correct (GGC) text is used. Experiments revealed that autoregressive ASRs show bias toward GGC text, leading to suboptimal performance. It is observed that there is an increase in performance in response to the decreased freedom of speech, with semi-spontaneous speech outperforming memorized speech. Aligner probabilities outperform confidence scores despite the alignment and overconfidence issues.

Abstract

Typically, in computer assisted language learning systems, pronunciation assessment of learners’ reading is a critical component when they’re practicing with long paragraphs. Thus, automatic methods need to be developed for assessing spoken paragraphs i.e. monologues. However, there is a limitation on the availability of monologues data, corresponding ratings and the works to assess them. To address the data and label scarcity, this work proposes a method to obtain vector (ProMoVec) representations to assess the pronunciation of monologue considering the i-vector modelling based on synthetic monologues data. The synthetic monologues are proposed to obtain by concatenating the utterances which are semantically similar and corresponding ratings predicted based on entropy criterion which uses utterance ratings. Experiments with SpeechOcean762 reveal that the PromoVecs used in downstream assessment task with clustering shows a relative improvement of 11.6% compared to the baseline method.

Abstract

Automatic pronunciation assessment is a critical component in computer assisted language learning. Typically, modeling pronunciation assessment tasks need labels, which are difficult to obtain as it requires expert annotators. Thus, it is essential to build an accurate model with less annotated data. In this work, an approach is proposed that considers a few speech samples using the i-vector framework. Each sample, first, is lengthened by T factor by concatenating the augmented samples of the same speech. The augmentation is obtained using time-scale modification (TSM), pitch-scale modification (PSM) and both. Next, phoneme-level goodness-of-pronunciation scores of concatenated speech are converted to a vector (GoP2Vec) with the i-vector framework. Experiments on two datasets revealed that the proposed GoP2Vec outperforms the state-of-the-art (SOTA) unsupervised methods and is on par with the SOTA supervised methods when it is used to train a simple neural model with a few samples.

Abstract

Computer-Assisted Language Learning systems provide speech exaggerating at mispronounced word locations as a feedback to the L2 learners. Traditionally, expert speakers recordings (which limit the scalability) are considered for this task though there are advancements in text-to-speech (TTS) that can generate native like natural sounding speech. To address these, this work proposes two novel controllable strategies for scalable speech exaggeration. One strategy is direct speech exaggeration that incorporates the proposed label conditioned tokenization in GlowTTS. Another strategy is cascading the state-of-the-art TTS to a WORLD vocoder with proposed energy and duration modifications. A subset of Tatoeba corpus, that we annotated with prominent words, is used for experimentation. Automatic and manual assessment reveals that the exaggerated speech quality from both direct and cascaded strategy with duration modification is closer to the prominent words in the native speaker's speech.

Abstract

Despite advancements in Speech-to-Speech Translation, maintaining expressiveness between the source and target speech, particularly between English and Indian languages, remains challenging. This study investigates prosodic similarities and variations between English and Malayalam, a language spoken in southern India. A set of 22 prompts from the IViE corpus, covering five categories—simple sentences, WH-questions, questions without morphosyntactic markers, inversion questions, and coordinations—was selected. These prompts, originally spoken by UK speakers, were translated into Malayalam and both language prompts were recorded by bilingual Malayalam speakers while preserving expressiveness. Word-level prominence was manually annotated, and comparisons were made across Indian English, Malayalam, and UK English. The analysis reveals that prominence is retained at key points in Indian English, whereas in Malayalam, it is low due to question specific diacritics and agglutinative nature.

Abstract

Word-level prominence plays a crucial role in spoken language understanding, as it helps to interpret the speaker’s intent. However, automatic word-level prominence detection remains underexplored, particularly in non-native speech, where prominence patterns are more variable due to native language influence. While prominence is determined by acoustic features such as energy, duration, and pitch, capturing statistics on these acoustics at the word level provides only a global representation, missing finer suprasegmental variations. In this study, we consider syllables as the suprasegmental unit and propose a methodology that jointly models syllable prominence variations and their contribution to word prominence using sequential neural networks. This enables learning word-level prominence representations with minimal labeled data. Our method outperforms an unsupervised n-gram-based baseline by 24.01% and a supervised SVM by 5.73%, demonstrating its effectiveness over both approaches.

Abstract

Automatic syllable stress detection is an essential component in Computer assisted language learning (CALL) systems to guide non-native language learners. In English, each word typically contains only one primary stressed syllable. However, standard loss functions, such as Binary Cross-Entropy (BCE), often result in predictions where multiple syllables may be stressed or none at all. As a result, automatic syllable stress detection models frequently require an additional post-processing step to ensure that only one syllable is stressed per word. This reliance on post-processing suggests that the model is not fully capturing the stress patterns accurately. To address this issue, we propose an adaptive weighted loss function that builds upon the Stress Intensity Modulation Loss proposed in our recent work of Post-Net. This adaptive weighted loss function is designed to enforce the constraint of a single primary stressed syllable directly during model training. We integrate this loss function into the previously proposed Post-Net (PN_DNN) and on a new architecture which is a hybrid of Post-Net and LSTM (PN_DLSTM). Their performance is compared against the state-of-the-art models trained with standard BCE loss. Experiments conducted on the ISLE corpus reveal that both the models trained only with BCE loss show a significant accuracy gap between with and without post-processing. In contrast, when these models are trained on the proposed adaptive weighted loss function, the gap is narrowed in both the models. Between the two models, the highest reduction is observed in PN_DNN with a decrease from 3.87% to 2.45% & 4.6% to 3.75% for GER & ITA respectively. This indicates that the adaptive weighted loss function effectively captures the linguistic constraint during training, reducing the need for post-processing.

Abstract

Automatic syllable stress detection is a crucial component in Computer-Assisted Language Learning (CALL) systems for language learners. Current stress detection models are typically trained on clean speech, which may not be robust in real-world scenarios where background noise is prevalent. To address this, speech enhancement (SE) models, designed to enhance speech by removing noise, might be employed, but their impact on preserving syllable stress patterns is not well studied. This study examines how different SE models, representing discriminative and generative modeling approaches, affect syllable stress detection under noisy conditions. We assess these models by applying them to speech data with varying signal-to-noise ratios (SNRs) from 0 to 20 dB, and evaluating their effectiveness in maintaining stress patterns. Additionally, we explore different feature sets to determine which ones are most effective for capturing stress patterns amidst noise. To further understand the impact of SE models, a human-based perceptual study is conducted to compare the perceived stress patterns in SE-enhanced speech with those in clean speech, providing insights into how well these models preserve syllable stress as perceived by listeners. Experiments are performed on English speech data from non-native speakers of German and Italian. And the results reveal that the stress detection performance is robust with the generative SE models when heuristic features are used. Also, the observations from the perceptual study are consistent with the stress detection outcomes under all SE models.

Abstract

Automatic detection of prominence at the word and syllable-levels is critical for building computer-assisted language learning systems. It has been shown that prosody embeddings learned by the current state-of-the-art (SOTA) text-to-speech (TTS) systems could generate word- and syllable-level prominence in the synthesized speech as natural as in native speech. To understand the effectiveness of prosody embeddings from TTS for prominence detection under nonnative context, a comparative analysis is conducted on the embeddings extracted from native and non-native speech considering the prominence-related embeddings: duration, energy, and pitch from a SOTA TTS named FastSpeech2. These embeddings are extracted under two conditions considering: 1) only text, 2) both speech and text. For the first condition, the embeddings are extracted directly from the TTS inference mode, whereas for the second condition, we propose to extract from the TTS under training mode. Experiments are conducted on native speech corpus: Tatoeba, and non-native speech corpus: ISLE. For experimentation, word-level prominence locations are manually annotated for both corpora. The highest relative improvement on word & syllable-level prominence detection accuracies with the TTS embeddings are found to be 13.7% & 5.9% and 16.2% & 6.9% compared to those with the heuristic-based features and self-supervised Wav2Vec-2.0 representations, respectively.