Renzheng Shi, M.Sc.

High-Fidelity Speech Coding with Neural Networks

Art der studentischen Arbeit: Masterarbeit

Betreuer: Renzheng Shi

Abteilung: Signalverarbeitung und Machine Learning

The master thesis is supervised by Renzheng Shi and Andreas Bär.

Communication is the key to information exchange and plays an important role in people’s daily life. Modern communication technologies, such as telephone calls, video chats, and social media software or streaming websites, dominate how we communicate with each other. Behind all the technologies, one essential component is data compression and coding achieved by audio and video codecs, which makes the transmission of information more convenient and faster.

Recently, neural network-based audio codecs like Lyra and Soundstream from Google, ENCODEC from Facebook, and Satin from Microsoft have been developed. Some of them have already been used to replace traditional coding software.

Are you interested in artificial intelligence and human-to-human/machine communication? Would you like to apply machine learning-based methods to improve our way of daily communication? We can find a close topic for you depending on your interests.

We focus on the coding of speech and audio signals. By taking advantage of the effectiveness of deep neural networks, a coding pipeline that optimizes the traditional coding methods is being built to give better quality after coding. In this work, possible topics are as follows:
- Networks that can synthesize high-quality speech signals would be studied
- Ablation study on how to improve the performance and reduce the complexity would be carried out
- Efficient coding methods that deliver the lossless coded speech and audio signal would be investigated
- ...

We aim at publishing the results from the thesis.

Successful attendance of the lectures ‘Mustererkennung’ and ‘Sprachkommunikation’ would be desirable but not necessary.

Learned speech and audio coding with variable bitrate

Art der studentischen Arbeit: Masterarbeit

Betreuer: Renzheng Shi

Abteilung: Signalverarbeitung und Machine Learning

Speech coding has benefited greatly from the advances of deep neural networks. A classic approach to accomplish this task is to make use of an autoencoder, consisting of an encoder-decoder structure and a quantizer.

The task of this thesis is to
- build up a model and adapt it to process speech features
- introduce entropy coding into this approach

Successful attendance of “Mustererkennung” and “Sprachkommunikation” would be desirable but not necessary.