Efficient, end-to-end and self-supervised methods for speech processing and generation
- Autores: Santiago Pascual de la Puente
- Directores de la Tesis: Antonio Bonafonte Cávez (dir. tes.), Joan Serra Julia (codir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2020
- Idioma: español
- Tribunal Calificador de la Tesis: Simon Kung (presid.), Marta Ruiz Costa-Jussà (secret.), Mireia Farrús i Cabeceran (voc.)
- Programa de doctorado: Programa de Doctorado en Teoría de la Señal y Comunicaciones por la Universidad Politécnica de Catalunya
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Deep learning has affected the speech processing and generation fields in many directions. First, end-to-end architectures allow the direct injection and synthesis of waveform samples. Secondly, the exploration of efficient solutions allow to implement these systems in computationally restricted environments, like smartphones. Finally, the latest trends exploit audio-visual data with least supervision. In this thesis these three directions are explored.
Firstly, we propose the use of recent pseudo-recurrent structures, like self-attention models and quasi-recurrent networks, to build acoustic models for text-to-speech. The proposed system, QLAD, turns out to synthesize faster on CPU and GPU than its recurrent counterpart whilst preserving the good synthesis quality level, which is competitive with state of the art vocoder-based models.
Then, a generative adversarial network is proposed for speech enhancement, named SEGAN. This model works as a speech-to-speech conversion system in time-domain, where a single inference operation is needed for all samples to operate through a fully convolutional structure. This implies an increment in modeling efficiency with respect to other existing models, which are auto-regressive and also work in time-domain. SEGAN achieves prominent results in noise supression and preservation of speech naturalness and intelligibility when compared to the other classic and deep regression based systems. We also show that SEGAN is efficient in transferring its operations to new languages and noises. A SEGAN trained for English performs similarly to this language on Catalan and Korean with only 24 seconds of adaptation data. Finally, we unveil the generative capacity of the model to recover signals from several distortions. We hence propose the concept of generalized speech enhancement. First, the model proofs to be effective to recover voiced speech from whispered one. Then the model is scaled up to solve other distortions that require a recomposition of damaged parts of the signal, like extending the bandwidth or recovering lost temporal sections, among others. The model improves by including additional acoustic losses in a multi-task setup to impose a relevant perceptual weighting on the generated result. Moreover, a two-step training schedule is also proposed to stabilize the adversarial training after the addition of such losses, and both components boost SEGAN's performance across distortions.Finally, we propose a problem-agnostic speech encoder, named PASE, together with the framework to train it. PASE is a fully convolutional network that yields compact representations from speech waveforms. These representations contain abstract information like the speaker identity, the prosodic features or the spoken contents. A self-supervised framework is also proposed to train this encoder, which suposes a new step towards unsupervised learning for speech processing. Once the encoder is trained, it can be exported to solve different tasks that require speech as input. We first explore the performance of PASE codes to solve speaker recognition, emotion recognition and speech recognition. PASE works competitively well compared to well-designed classic features in these tasks, specially after some supervised adaptation. Finally, PASE also provides good descriptors of identity for multi-speaker modeling in text-to-speech, which is advantageous to model novel identities without retraining the model.
