|
| |
|
|
Modelování dynamiky prosodie pro rozpoznávání řečníka
Jančík, Zdeněk ; Fapšo, Michal (oponent) ; Matějka, Pavel (vedoucí práce)
V současných systémech pro rozpoznání mluvčího se zpravidla využívají krátkodobé akustické příznaky. Jiné příznaky se používají jen zřídka. V práci se zaměřím na prosodické příznaky získané z průběhu základního tónu a energie. Tyto příznaky modelují průběh základního tónu v jednotlivých fonémech nebo slabikách. Z literatury je známo, že systémy založené na prosodii neposkytují tak dobré výsledky jako akustické, ale spojením akustického systému a systému založeného na prosodii se dosáhne značného zlepšení výsledků. To ověřím spojením s akustickým systémem vyvinutým na VUT. Při experimentech použiji data z evaluací pořádaných Národním úřadem pro standardy a technologie (NIST).
|
|
|
Microphone Arrays for Speaker Recognition
Mošner, Ladislav ; Plchot, Oldřich (oponent) ; Černocký, Jan (vedoucí práce)
This thesis addresses the problem of remote speaker recognition. The accuracy of standard speaker recognition decreases considerably in the presence of far-field data, therefore, we devised two strategies to improve the results. First, we employed a microphone array (purposely positioned set of microphones) that is able to steer a virtual "beam" to the position of the speaker. We also performed system adaptation of different parts of the system (PLDA scoring and i-vector extraction). We have synthesized our training and test data from the standard NIST 2010 data by room simulation and we have shown that both techniques and their combination significantly improve the results. We have also dealt with joint speaker identity and position estimation. While the results in simulated outdoor environment (reverberation-free) are encouraging, the results from interiors (with reverberation) are mixed and require further investigation. Finally, we were able to test our system on a limited amount of real re-transmitted data. While the results for male speakers match the simulation, the results for females are not convincing and need further analysis.
|
|
|
Resilience of Biometric Authentication of Voice Assistants against Deepfakes
Šandor, Oskar ; Firc, Anton (oponent) ; Malinka, Kamil (vedoucí práce)
With the rise of deepfake technology, imitating the voice of strangers has become a lot easier. It is no longer necessary to have a professional impersonator to imitate the voice of a person to possibly deceive a human or machine. Attackers only need a few recordings of a person's voice, regardless of the content, to create a voice clone using online or open-source tools. In that case, he or she can create recordings with content that the person may have never said. These recordings can be misused, for example, for unauthorized use of voice-assistant devices. The aim of this work is to determine whether voice assistants can recognize synthetized recordings (deepfakes). Experiments conducted in this thesis show that deepfakes created in a matter of minutes can spoof speaker recognition in voice assistants and can be used to carry out several attacks.
|
|
|
Hledání nových cest v rozpoznávání řečníka založeného na neuronových sítích
Sova, Damián ; Matějka, Pavel (oponent) ; Glembek, Ondřej (vedoucí práce)
Keďže zadanie tejto práce je veľmi široké, tak sa bolo treba sústrediť len na určitú sféru. Nakoniec, cieľom tejto práce je aplikovať optimalizačnú metódu Stochastického Spriemerovania Váh do tréningového procesu Hlbokej Neurónovej Siete. Po predstavení potrebných teoretických vedomostí v prvej časti práce, nasleduje druhá časť s priebehmi jednotlivých experimentov. V teoretickej časti je dôraz kladený hlavne na objasnenie celého životného cyklu trénovacieho a vyhodnocovacieho procesu, vrátane popisu jednotlivých komponentov. Praktická časť poskytuje podrobný pohľad na každý experiment, ktorých cieľom je demonštrovať dosiahnuteľnosť zvýšenia výkonnosti systému rozpoznávania rečníka. Celkové zlepšenie výkonu sa podarilo dosiahnuť postupným aplikovaním rôznych tréningových konfigurácií, v ktorých sa zohľadňujú skúsenosti z predchádzajúcich experimentov. Kľúčovou zložkou úspešného Stochastického Spriemerovania Váh v experimentoch bola dostatočne vysoká konštantná hodnota Miery Učenia s aplikovaným postupným prechodom alebo Cyklický priebeh Miery Učenia.
|
|
|
Unsupervised Evaluation of Speaker Recognition System
Odehnal, Ondřej ; Plchot, Oldřich (oponent) ; Matějka, Pavel (vedoucí práce)
The context of this thesis is the state-of-the-art system for speaker identification (SID) based on the deep nerual network with x-vector embeddings. This thesis aims to propose and experimentally assess several techniques for evaluating the SID system using unlabelled datasets. For this purpose, discriminative embedding is created for every recording in the dataset. These embeddings are used to cluster the recordings and thus create pseudo-labels corresponding to different clusters. The SID system evaluation is based on equal error rate (EER), which uses these pseudo-labels. We proposed several unsupervised learning algorithms to achieve this; K-means, Gaussian mixture models (GMM), and agglomerative hierarchical clustering (AHC). After thorough testing, the K-means model with the Silhouette value showed the best results. This method achieved an estimate of 5.72 % EER with the reference EER equal to 5.15 % on SITW dev-core-core. Similar results were observed on the SITW eval-core-core, where the estimated EER is equal to 5.86 % and the reference 5.08 %. The difference between estimated and reference EER is 0.57 % for the dev-core-core and 0.78 % for the eval-core-core. Another series of experiments were conducted on NIST SRE16 and VoxCeleb1 to verify robustness of the proposed method. Generally, the developed testing process had an estimated error of around 1 % in all test databases, an excellent result for an unsupervised learning technique.
|
|
|
Personal Voice Activity Detection
Sedláček, Šimon ; Landini, Federico Nicolás (oponent) ; Švec, Ján (vedoucí práce)
This work aims to implement, test, and evaluate a speaker-conditioned Voice Activity Detection (VAD) method called Personal VAD. The method builds upon an LSTM-based approach to VAD and its purpose is to introduce a system that can reliably detect speech of a target speaker, while retaining the typical characteristics of a VAD system, mainly in terms of small model size, low latency, and low necessary computational resources. The system is trained to distinguish between three classes: non-speech, target speaker speech, and non-target speaker speech. For this purpose, the method utilizes speaker embeddings as a part of the input feature vector to represent the target speaker. Some of the more heavyweight personal VAD variants also make use of speaker verification scores issued to each frame based on the target embedding, resulting in a more robust system. In addition to the one scoring method presented in the original article, two other scoring approaches are introduced, both outperforming the baseline method and improving the performance even for acoustically challenging conditions.
|
|
|
Robustní rozpoznávání mluvčího pomocí neuronových sítí
Profant, Ján ; Rohdin, Johan Andréas (oponent) ; Matějka, Pavel (vedoucí práce)
Tématem této práce je analýza nejmodernějších systémů pro rozpoznávání řečníka za použití neurónových sítí (nazývaných x-vektory) v rozličných podmínkách, jako jsou širokopásmové a úzkopásmové data, který je robustní vůči neviděnému jazyku, specifickému hluku nebo telefonnimu kodeku. Automatický systém mapuje zvukovou nahrávku variabilní délky do fixně dlouhého vektoru, který je následně využit jako reprezentace řečníka. V této práci jsme porovnali systémy založené na neurónových sítich s výsledkem VUT týmu v Speakers in the Wild Speaker Recognition Challenge (SITW), který využíval donedávna velmi populární statistický model - i-vektory. Pozorovali jsme, že s nedávno publikovanými x-vektory dosahujeme 4.38 krát nižší Equal Error Rate pro SITW core-core evaluační sadu v porovnání s výsledkem z roku 2016 od VUT v SITW soutěži. Kromě toho jsme ukázali, že diarizace v nahrávkach s více mluvčími významně snižuje chybovost systému pro SITW core-multi evaluační data, ale podobný trend jsme neviděli pro dataset NIST SRE 2018 VAST.
|
|
|
Microphone Arrays for Speaker Recognition
Mošner, Ladislav ; Plchot, Oldřich (oponent) ; Černocký, Jan (vedoucí práce)
This thesis addresses the problem of remote speaker recognition. The accuracy of standard speaker recognition decreases considerably in the presence of far-field data, therefore, we devised two strategies to improve the results. First, we employed a microphone array (purposely positioned set of microphones) that is able to steer a virtual "beam" to the position of the speaker. We also performed system adaptation of different parts of the system (PLDA scoring and i-vector extraction). We have synthesized our training and test data from the standard NIST 2010 data by room simulation and we have shown that both techniques and their combination significantly improve the results. We have also dealt with joint speaker identity and position estimation. While the results in simulated outdoor environment (reverberation-free) are encouraging, the results from interiors (with reverberation) are mixed and require further investigation. Finally, we were able to test our system on a limited amount of real re-transmitted data. While the results for male speakers match the simulation, the results for females are not convincing and need further analysis.
|
|
|
Optimization of Gaussian Mixture Subspace Models and Related Scoring Algorithms in Speaker Verification
Glembek, Ondřej ; Brummer, Niko (oponent) ; Campbell,, William (oponent) ; Burget, Lukáš (vedoucí práce)
This thesis deals with Gaussian Mixture Subspace Modeling in automatic speaker recognition. The thesis consists of three parts. In the first part, Joint Factor Analysis (JFA) scoring methods are studied. The methods differ mainly in how they deal with the channel of the tested utterance. The general JFA likelihood function is investigated and the methods are compared both in terms of accuracy and speed. It was found that linear approximation of the log-likelihood function gives comparable results to the full log-likelihood evaluation while simplyfing the formula and dramatically reducing the computation speed. In the second part, i-vector extraction is studied and two simplification methods are proposed. The motivation for this part was to allow for using the state-of-the-art technique on small scale devices and to setup a simple discriminative-training system. It is shown that, for long utterances, while sacrificing the accuracy, we can get very fast and compact i-vector systems. On a short-utterance(5-second) task, the results of the simplified systems are comparable to the full i-vector extraction. The third part deals with discriminative training in automatic speaker recognition. Previous work in the field is summarized and---based on the knowledge from the earlier chapters of this work---discriminative training of the i-vector extractor parameters is proposed. It is shown that discriminative re-training of the i-vector extractor can improve the system if the initial estimation is computed using the generative approach.
|
host ::
RSS.