Title: A Comparison of Speaker-based and Utterance-based Data Selection for Text-to-Speech Synthesis
Building on previous work in subset selection of training data for text-to-speech (TTS), this work compares speaker-level and utterance-level selection of TTS training data, using acoustic features to guide selection. We find that speaker-based selection is more effective than utterance-based selection, regardless of whether selection is guided by a single feature or a combination of features. We use US English telephone data collected for automatic speech recognition to simulate the conditions of TTS training on low-resource languages. Our best voice achieves a human-evaluated WER of 29.0% on semantically-unpredictable sentences. This constitutes a significant improvement over our baseline voice trained on the same amount of randomly selected utterances, which performed at 42.4% WER. In addition to subjective voice evaluations with Amazon Mechanical Turk, we also explored objective voice evaluation using mel-cepstral distortion. We found that this measure correlates strongly with human evaluations of intelligibility, indicating that it may be a useful method to evaluate or pre-select voices in future work. more »« less
Cohn, Michelle; Zellou, Georgia(
, Proceedings of Interspeech)
null
(Ed.)
This study tests speech-in-noise perception and social ratings of speech produced by different text-to-speech (TTS) synthesis methods. We used identical speaker training datasets for a set of 4 voices (using AWS Polly TTS), generated using neural and concatenative TTS. In Experiment 1, listeners identified target words in semantically predictable and unpredictable sentences in concatenative and neural TTS at two noise levels (-3 dB, -6 dB SNR). Correct word identification was lower for neural TTS than for concatenative TTS, in the lower SNR, and for semantically unpredictable
sentences. In Experiment 2, listeners rated the voices on 4 social attributes. Neural TTS was rated as more human-like, natural, likeable, and familiar than concatenative TTS. Furthermore, how natural listeners rated the neural TTS voice was positively related to their speech-in-noise accuracy. Together, these findings show that the TTS method influences both intelligibility and social judgments of speech — and that these patterns are linked. Overall, this work contributes to our understanding of the of the nexus of speech technology and human speech perception.
Alemu, Yared; Chen, Hua; Duan, Chenghao; Caulley, Desmond; Arriaga, Rosa I; Sezgin, Emre(
, JMIR Research Protocols)
Background Even before the onset of the COVID-19 pandemic, children and adolescents were experiencing a mental health crisis, partly due to a lack of quality mental health services. The rate of suicide for Black youth has increased by 80%. By 2025, the health care system will be short of 225,000 therapists, further exacerbating the current crisis. Therefore, it is of utmost importance for providers, schools, youth mental health, and pediatric medical providers to integrate innovation in digital mental health to identify problems proactively and rapidly for effective collaboration with other health care providers. Such approaches can help identify robust, reproducible, and generalizable predictors and digital biomarkers of treatment response in psychiatry. Among the multitude of digital innovations to identify a biomarker for psychiatric diseases currently, as part of the macrolevel digital health transformation, speech stands out as an attractive candidate with features such as affordability, noninvasive, and nonintrusive. Objective The protocol aims to develop speech-emotion recognition algorithms leveraging artificial intelligence/machine learning, which can establish a link between trauma, stress, and voice types, including disrupting speech-based characteristics, and detect clinically relevant emotional distress and functional impairments in children and adolescents. Methods Informed by theoretical foundations (the Theory of Psychological Trauma Biomarkers and Archetypal Voice Categories), we developed our methodology to focus on 5 emotions: anger, happiness, fear, neutral, and sadness. Participants will be recruited from 2 local mental health centers that serve urban youths. Speech samples, along with responses to the Symptom and Functioning Severity Scale, Patient Health Questionnaire 9, and Adverse Childhood Experiences scales, will be collected using an Android mobile app. Our model development pipeline is informed by Gaussian mixture model (GMM), recurrent neural network, and long short-term memory. Results We tested our model with a public data set. The GMM with 128 clusters showed an evenly distributed accuracy across all 5 emotions. Using utterance-level features, GMM achieved an accuracy of 79.15% overall, while frame selection increased accuracy to 85.35%. This demonstrates that GMM is a robust model for emotion classification of all 5 emotions and that emotion frame selection enhances accuracy, which is significant for scientific evaluation. Recruitment and data collection for the study were initiated in August 2021 and are currently underway. The study results are likely to be available and published in 2024. Conclusions This study contributes to the literature as it addresses the need for speech-focused digital health tools to detect clinically relevant emotional distress and functional impairments in children and adolescents. The preliminary results show that our algorithm has the potential to improve outcomes. The findings will contribute to the broader digital health transformation. International Registered Report Identifier (IRRID) DERR1-10.2196/46970
Khan, Awais; Malik, Khalid Mahmood(
, MAD '23: Proceedings of the 2nd ACM International Workshop on Multimedia AI against Disinformation)
The prevalence of voice spoofing attacks in today’s digital world has become a critical security concern. Attackers employ various techniques, such as voice conversion (VC) and text-to-speech (TTS), to generate synthetic speech that imitates the victim’s voice and gain access to sensitive information. The recent advances in synthetic speech generation pose a significant threat to modern security systems, while traditional voice authentication methods are incapable of detecting them effectively. To address this issue, a novel solution for logical access (LA)-based synthetic speech detection is proposed in this paper. SpoTNet is an attention-based spoofing transformer network that includes crafted front-end spoofing features and deep attentive features retrieved using the developed logical spoofing transformer encoder (LSTE). The derived attentive features were then processed by the proposed multi-layer spoofing classifier to classify speech samples as bona fide or synthetic. In synthetic speeches produced by the TTS algorithm, the spectral characteristics of the synthetic speech are altered to match the target speaker’s formant frequencies, while in VC attacks, the temporal alignment of the speech segments is manipulated to preserve the target speaker’s prosodic features. By highlighting these observations, this paper targets the prosodic and phonetic-based crafted features, i.e., the Mel-spectrogram, spectral contrast, and spectral envelope, presenting an effective preprocessing pipeline proven to be effective in synthetic speech detection. The proposed solution achieved state-of-the-art performance against eight recent feature fusion methods with lower EER of 0.95% on the ASVspoof-LA dataset, demonstrating its potential to advance the field of speaker identification and improve speaker recognition systems.
With the advent of automated speaker verifcation (ASV) systems comes an equal and
opposite development: malicious actors may seek to use voice spoofng attacks to fool
those same systems. Various counter measures have been proposed to detect these spoofing attacks, but current oferings in this arena fall short of a unifed and generalized
approach applicable in real-world scenarios. For this reason, defensive measures for ASV
systems produced in the last 6-7 years need to be classifed, and qualitative and quantitative comparisons of state-of-the-art (SOTA) counter measures should be performed to
assess the efectiveness of these systems against real-world attacks. Hence, in this work,
we conduct a review of the literature on spoofng detection using hand-crafted features,
deep learning, and end-to-end spoofng countermeasure solutions to detect logical access
attacks, such as speech synthesis and voice conversion, and physical access attacks, i.e.,
replay attacks. Additionally, we review integrated and unifed solutions to voice spoofng
evaluation and speaker verifcation, and adversarial and anti-forensic attacks on both voice
counter measures and ASV systems. In an extensive experimental analysis, the limitations
and challenges of existing spoofng counter measures are presented, the performance of
these counter measures on several datasets is reported, and cross-corpus evaluations are
performed, something that is nearly absent in the existing literature, in order to assess the
generalizability of existing solutions. For the experiments, we employ the ASVspoof2019,
ASVspoof2021, and VSDC datasets along with GMM, SVM, CNN, and CNN-GRU classifers. For reproducibility of the results, the code of the testbed can be found at our GitHub
Repository (https://github.com/smileslab/Comparative-Analysis-Voice-Spoofing).
This paper describes experiments in training HMM-based text-to-speech (TTS) voices on data collected for Automatic Speech Recognition (ASR) training. We compare a number of filtering techniques designed to identify the best utterances from a noisy, multi-speaker corpus for training voices, to exclude speech containing noise and to include speech close in nature to more traditionally-collected TTS corpora. We also evaluate the use of automatic speech recognizers for intelligibility assessment in comparison with crowdsourcing methods. While the goal of this work is to develop natural-sounding and intelligible TTS voices in Low Resource Languages (LRLs) rapidly and easily, without the expense of recording data specifically for this purpose, we focus on English initially to identify the best filtering techniques and evaluation methods. We find that, when a large amount of data is available, selecting from the corpus based on criteria such as standard deviation of f0, fast speaking rate, and hypo-articulation produces the most intelligible voices.
Kai-Zhan Lee, Erica Cooper. A Comparison of Speaker-based and Utterance-based Data Selection for Text-to-Speech Synthesis. Retrieved from https://par.nsf.gov/biblio/10097223. Interspeech 2018 12873-2877. Web. doi:DOI: 10.21437/Interspeech.2018-1313.
Kai-Zhan Lee, Erica Cooper. A Comparison of Speaker-based and Utterance-based Data Selection for Text-to-Speech Synthesis. Interspeech 2018, 12873-2877 (). Retrieved from https://par.nsf.gov/biblio/10097223. https://doi.org/DOI: 10.21437/Interspeech.2018-1313
@article{osti_10097223,
place = {Country unknown/Code not available},
title = {A Comparison of Speaker-based and Utterance-based Data Selection for Text-to-Speech Synthesis},
url = {https://par.nsf.gov/biblio/10097223},
DOI = {DOI: 10.21437/Interspeech.2018-1313},
abstractNote = {Building on previous work in subset selection of training data for text-to-speech (TTS), this work compares speaker-level and utterance-level selection of TTS training data, using acoustic features to guide selection. We find that speaker-based selection is more effective than utterance-based selection, regardless of whether selection is guided by a single feature or a combination of features. We use US English telephone data collected for automatic speech recognition to simulate the conditions of TTS training on low-resource languages. Our best voice achieves a human-evaluated WER of 29.0% on semantically-unpredictable sentences. This constitutes a significant improvement over our baseline voice trained on the same amount of randomly selected utterances, which performed at 42.4% WER. In addition to subjective voice evaluations with Amazon Mechanical Turk, we also explored objective voice evaluation using mel-cepstral distortion. We found that this measure correlates strongly with human evaluations of intelligibility, indicating that it may be a useful method to evaluate or pre-select voices in future work.},
journal = {Interspeech 2018},
volume = {12873-2877},
author = {Kai-Zhan Lee, Erica Cooper},
}
Warning: Leaving National Science Foundation Website
You are now leaving the National Science Foundation website to go to a non-government website.
Website:
NSF takes no responsibility for and exercises no control over the views expressed or the accuracy of
the information contained on this site. Also be aware that NSF's privacy policy does not apply to this site.