|
|
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.
|
|
|
Non-Parallel Voice Conversion
Brukner, Jan ; Plchot, Oldřich (oponent) ; Černocký, Jan (vedoucí práce)
Voice conversion (VC) aims at converting the voice of source speaker to the voice of target speaker. It is popular in funny Internet videos but has also series of serious use cases, such as dubbing of audiovisual material and anonymization of voice (for example for witness protection). As it can serve for spoofing of voice identification systems, it is also an important tool for development spoofing detectors and counter-measures. Training VC models has mainly been on parallel audios (ie. two speakers uttering the same text) and on high quality audio material. The goal of this thesis was to investigate developing VC on non-parallel data and with low quality signals, mainly from publicly available dataset VoxCeleb. This work follows the state-of-the-art AutoVC architecture defined by Qian et al. It is based on neural network (NN) autoencoders, aiming to separate speech into content- and speaker-dependent embedding. The target speech is then obtained by replacing source speaker embedding by the target speaker one. We have improved Qian's architecture to process low-quality audio by experimenting with different speaker embeddings (d-vectors vs. x-vectors), introducing a speaker classifier from content embeddings in an adversarial setup, and tuning the size of content embeddings imposing an information bottleneck to the autoencoder. Also, we have defined another adversarial architecture by comparing original content embeddings with those obtained after the VC process. The results of experiments prove that non-parallel VC on low-quality data is indeed doable. The resulting audios were not so good as in case of using high-quality ones, but the speaker verification results after spoofing by proposed system have clearly shown a shift of voice characteristics toward the target speakers.
|
|
|
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.
|
|
|
Non-Parallel Voice Conversion
Brukner, Jan ; Plchot, Oldřich (oponent) ; Černocký, Jan (vedoucí práce)
Voice conversion (VC) aims at converting the voice of source speaker to the voice of target speaker. It is popular in funny Internet videos but has also series of serious use cases, such as dubbing of audiovisual material and anonymization of voice (for example for witness protection). As it can serve for spoofing of voice identification systems, it is also an important tool for development spoofing detectors and counter-measures. Training VC models has mainly been on parallel audios (ie. two speakers uttering the same text) and on high quality audio material. The goal of this thesis was to investigate developing VC on non-parallel data and with low quality signals, mainly from publicly available dataset VoxCeleb. This work follows the state-of-the-art AutoVC architecture defined by Qian et al. It is based on neural network (NN) autoencoders, aiming to separate speech into content- and speaker-dependent embedding. The target speech is then obtained by replacing source speaker embedding by the target speaker one. We have improved Qian's architecture to process low-quality audio by experimenting with different speaker embeddings (d-vectors vs. x-vectors), introducing a speaker classifier from content embeddings in an adversarial setup, and tuning the size of content embeddings imposing an information bottleneck to the autoencoder. Also, we have defined another adversarial architecture by comparing original content embeddings with those obtained after the VC process. The results of experiments prove that non-parallel VC on low-quality data is indeed doable. The resulting audios were not so good as in case of using high-quality ones, but the speaker verification results after spoofing by proposed system have clearly shown a shift of voice characteristics toward the target speakers.
|
host ::
RSS.