Low-bit-rate Informed Source Separation Using Decoder NTF and Efficient Parameter Coding

Informed Source Separation (ISS) unifies the fields of audio source separation and audio coding: ISS uses source separation methods for efficient coding of audio objects (e.g. recordings of musical instruments). The basic idea is that the source separation step is supported by a compact set of side information which is extracted at the encoder side given the original recordings. The resulting side information is then transmitted to the decoder which performs source separation from the mixture.

This thesis deals with ISS algorithms compressing the audio objects with nonnegative factorization methods such as Nonnegative Tensor Factorization (NTF). These methods are widely used in the source separation community as they allow for an efficient description of single sound events present in audio recordings. The novel contributions are as follows:

Context-based Adaptive Binary Arithmetic Coding (CABAC) is adapted to the field of factorization-based ISS for coding the quantized NTF parameters by suitable context models. Experimental results show that CABAC outperforms other existing entropy coding schemes.

The decoder is extended to use a complete NTF-based blind source separation algorithm. It is shown experimentally that this extension enables very low bit rates.

To correct possible errors of the source separation step in the decoder, the encoder can compute residuals in time-frequency domain. The thesis proposes to quantize these residuals under a rate-distortion constraint with Rate-distortion Optimized Quantization (RDOQ). In experiments, the proposed method is compared to other state-of-the-art coding which combines ISS with source coding.