Using Deep Neural Networks for Identification of Slavic Languages from Acoustic Signal

Mateju, Lukas; Červa, Petr; Žďánský, Jindřich; Safarik, Radek

doi:10.21437/interspeech.2018-1165

Cited by 12 publications

(19 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Applying ASR methods to SLI, e.g. by training language classifiers on phoneme embeddings extracted from a phoneme recognizer, has shown to work very well [19,20,21,22]. While end-to-end SLI performed directly on labeled speech features is usually outperformed by models that utilize phoneme level information, it is sometimes possible to reach good performance also with end-to-end models [6,23].…”

Section: End-to-end Deep Learning Sli Toolkitmentioning

confidence: 99%

See 1 more Smart Citation

Releasing a Toolkit and Comparing the Performance of Language Embeddings Across Various Spoken Language Identification Datasets

Lindgren¹,

Jauhiainen²,

Kurimo³

2020

Interspeech 2020

View full text Add to dashboard Cite

In this paper, we propose a software toolkit for easier end-toend training of deep learning based spoken language identification models across several speech datasets. We apply our toolkit to implement three baseline models, one speaker recognition model, and three x-vector architecture variations, which are trained on three datasets previously used in spoken language identification experiments. All models are trained separately on each dataset (closed task) and on a combination of all datasets (open task), after which we compare if the open task training yields better language embeddings. We begin by training all models end-to-end as discriminative classifiers of spectral features, labeled by language. Then, we extract language embedding vectors from the trained end-to-end models, train separate Gaussian Naive Bayes classifiers on the vectors, and compare which model provides best language embeddings for the backend classifier. Our experiments show that the open task condition leads to improved language identification performance on only one of the datasets. In addition, we discovered that increasing x-vector model robustness with random frequency channel dropout significantly reduces its end-to-end classification performance on the test set, while not affecting back-end classification performance of its embeddings. Finally, we note that two baseline models consistently outperformed all other models.

show abstract

Section: End-to-end Deep Learning Sli Toolkitmentioning

confidence: 99%

“…The test set contains test utterances of varying length, with the median duration at 15 seconds. DoSL Dataset of Slavic Languages (DoSL), contains speech in 11 Slavic languages [19]. The dataset includes 220 hours of training data and 8 hours of test data, where test utterances are almost uniformly distributed between 5 and 6 seconds.…”

Section: Datasetsmentioning

confidence: 99%

Releasing a Toolkit and Comparing the Performance of Language Embeddings Across Various Spoken Language Identification Datasets

Lindgren¹,

Jauhiainen²,

Kurimo³

2020

Interspeech 2020

View full text Add to dashboard Cite

show abstract

“…In addition to the 39-dimensional MFCCs, we have also utilized 13-dimensional MFCCs with ∆ and ∆∆ coefficients (i.e., a 39-dimensional feature vector as well), and 39-dimensional bottleneck features (BTNs) extracted from the DNN trained for speech recognition (as suggested for the speaker and language identification, e.g., in [39]). Detailed information about our BTN feature extractor can be found in [40].…”

Section: Acoustic Featuresmentioning

confidence: 99%

An Approach to Online Speaker Change Point Detection Using DNNs and WFSTs

2019

Self Cite

View full text Add to dashboard Cite

In this paper, a new approach to speaker change point (SCP) detection is presented. This method is suitable for online applications (e.g., real-time broadcast monitoring). It is designed in a series of consecutive experiments, aiming at quality of detection as well as low latency. The resulting scheme utilizes a convolution neural network (CNN), whose output is smoothed by a decoder. The CNN is trained using data complemented by artificial examples to reduce different types of errors, and the decoder is based on a weighted finite state transducer (WFST) with the forced length of the transition model. Results obtained on data taken from the COST278 database show that our online approach yields results comparable with an offline multi-pass LIUM toolkit while operating online with a low latency.

show abstract

“…Although DNF-based systems provide the best LID performance, the ASR must be trained first, which requires a large volume of phoneme labeling data. Recently, end-to-end approaches with recurrent neural networks (RNN), convolutional neural networks, and attention-based neural networks have been investigated on LID tasks [8,9].…”

Section: Introductionmentioning

confidence: 99%

“…In this paper, we focus on the identification of variable length utterance spoken language from webcast using LSTM-based and self-attention [11] CNN end-to-end approaches [9,10]. We explore the LID models based on global average pooling.…”

Section: Introductionmentioning

confidence: 99%

Identification of Spoken Language from Webcast Using Deep Convolutional Recurrent Neural Networks

Zhu¹,

Huang²,

Yang³

et al. 2019

dtcse

View full text Add to dashboard Cite

This paper investigated two end-to-end approaches for the identification of spoken language from webcast sources. Long short-term memory (LSTM) and self-attention mechanism architectures are adopted and compared against a deep convolution network baseline model. These methods focused on the performance of spoken language identification (LID) on variable length utterance. The dataset used for experimental evaluation contains five language data collected from webcast (Webcast-5) and ten Chinese dialect language datasets from IFLYTEK (IFLYTEK-10). The end-to-end LID system was trained using five kinds of acoustic features: Mel-frequency cepstral coefficients (MFCCs), shifted delta cepstral coefficients (SDC), Perceptual Linear Prediction (PLP), log Mel-scale filter bank energies (Fbank) and spectrogram energies. The best model using a single feature set achieves an accuracy of 79.6% and a C avg of 10.87%.

show abstract

Using Deep Neural Networks for Identification of Slavic Languages from Acoustic Signal

Cited by 12 publications

References 23 publications

Releasing a Toolkit and Comparing the Performance of Language Embeddings Across Various Spoken Language Identification Datasets

Releasing a Toolkit and Comparing the Performance of Language Embeddings Across Various Spoken Language Identification Datasets

An Approach to Online Speaker Change Point Detection Using DNNs and WFSTs

Identification of Spoken Language from Webcast Using Deep Convolutional Recurrent Neural Networks

Contact Info

Product

Resources

About