Frame-synchronized blind speech watermarking via improved adaptive mean modulation and perceptual-based additive modulation in DWT domain

Highlights

• A DWT-based algorithm is introduced to perform efficient blind speech watermarking.

• Schemes for embedding synchronization codes and information bits are different.

• Perceptual additive modulation renders imperceptible and reliable detection.

• Improved adaptive mean modulation secures an accurate retrieval of watermark bits.

• The proposed algorithm outperforms three others with capacity set at 200 bps.

Abstract

This paper presents a blind speech watermarking algorithm that adopts different strategies to embed synchronization codes and information bits into separate DWT subbands. Speech frames for data hiding are chosen based on intensity thresholding. Under the guidance of the source-filter theory, a bipolar synchronization code sequence is substituted for the noisy part of the inverse filtered excitation in the 2nd level detail subband. Such a design allows the increase of embedding strength without adversely disturbing the pitch harmonics of voiced speech. Experiment results show that a synchronization code of size 640 is sufficient to render reliable detection. The PESQ metrics indicate that the quality degradation due to the synchronization code embedding is almost negligible. As for the binary embedding in the 2nd level approximation subband, we improve the adaptive mean modulation scheme to secure the retrieval of watermark bits on a frame basis. Experiment results confirm that, with the payload capacity set at 200 bit per second (bps), the proposed scheme demonstrates better robustness than three other DWT-based methods in the presence of commonly encountered signal processing attacks. Furthermore, reducing the capacity down to 100 bps alters the low-frequency spectral distribution, which contributes noteworthy improvements in robustness and imperceptibility.

Introduction

With the rapid development of network and information technologies, nowadays people can reproduce and disseminate digital multimedia data throughout the world much easier than ever before. However, the illegal use of multimedia data is also rampant in the current era. The copyright infringement problem has gradually become a serious issue to be solved [22]. Digital watermarking technology is a promising technique, widely used for several purposes including content authentication, ownership verification, covert communication, broadcast monitoring and fingerprinting, etc. It has become a hot topic in the field of communication and information security in recent years [6], [12], [22], [32].

Watermarking technology generally takes into consideration of four factors: imperceptibility, security, robustness, and capacity [6], [12]. An ideal watermarking algorithm is expected to hide a sufficient amount of information into the host signal in an imperceptible manner but still capable of resisting against malicious attacks. There are several ways to classify watermarking techniques. One such a classification is “robust” versus “fragile”. Robust watermarks are strongly resistant to attacks, whereas fragile watermarks are supposed to crumble under any attempt of tampering. Furthermore, depending on the need of the original host signal in the extraction process, watermarking schemes can be categorized as blind, semi-blind, and non-blind. Blind watermarking is designed to recover an embedded watermark without the participation of the original signal, while the non-blind approach can only be carried out using the original source. Semi-blind watermarking involves situations where information other than the source itself is needed for watermark extraction.

Over the past two decades, numerous watermarking methods have been developed for multimedia data such as images, audios and videos. Nonetheless, far less attention was paid to the watermarking of speech signals. Speech is a specific type of audio signal; therefore, the techniques developed for audio watermarking are potentially applicable to speech watermarking. However, speech differs from typical audios in aspects including the temporal continuity, spectral intensity distribution, production modeling, and processing scenarios [24], [30], [33]. The techniques developed for audio watermarking may not be suitable for speech watermarking [32].

Among the watermarking methods dedicated to speech signals, the one developed by Hofbauer et al. [14] exploited the fact that human ears are insensitive to the phase of non-voiced speech. Their method was focused on replacing the excitation signal of an autoregressive representation in non-voiced segments. On the other hand, Coumou and Sharma [5] embedded data via pitch modification in voiced segments. To avoid insertion, deletion, and substitution errors in the estimates of embedded data, they also resorted to a concatenated coding scheme to safeguard synchronization and error recovery. In another watermarking scheme developed based on the codebook-excited linear prediction (CELP) codec, Chen and Liu [3] modified the position indices of selected excitation pulses.

The exploitation of the spectral envelope of the speech signal has also been attempted. Several methods emerged from the linear prediction (LP) analysis of speech. In [30], Faundez-Zanuy et al. hid the watermark information below the formant peaks of the LP spectrum. Chen and Zhu [2] achieved robust watermarking by inserting watermark bits inside codebook indices, while applying multistage vector quantization (MSVQ) to the derived LP coefficients. Yan and Guo [43] converted the LP coefficients to inverse sine (IS) parameters. Watermark embedding was achieved by manipulating the IS parameters using odd–even modulation [23].

The implementation of speech watermarking can certainly take reference from existing audio watermarking methods. In order to take advantage of signal characteristics and/or auditory properties [22], many researchers tended to conduct the watermarking process in transform domains such as discrete Fourier transform (DFT) [7], [27], [31], [36], discrete cosine transform (DCT) [16], [25], [26], [39], [44], discrete wavelet transform (DWT) [1], [4], [18], [38], [39], [40], [41], and cepstrum [15], [28], [29]. Among the transforms used to perform audio watermarking, DWT is currently the most popular due to its perfect reconstruction and good multi-resolution properties. The effectiveness of the DWT-based approach in audio watermarking suggests that the same technique may work for speech watermarking as well if speech characteristics can be adequately taken into account.

Lei et al. [24] proposed a sophisticated wavelet-based scheme especially for breath sounds. Their scheme required the use of lifting wavelet transform (LWT), discrete cosine transform (DCT), singular value decomposition (SVD) along with the use of particle swarm optimization (PSO) technique to optimize the quantization steps for dither modulation (DM). Nematollahi et al. [33] embedded watermark bits by quantizing the eigenvalue derived from the singular value decomposition of the approximation coefficients in the DWT domain. Hu et al. [19] took advantage of the multi-resolution analysis properties of DWT, which eventually led to the development of a synchronous package scheme. In their design, watermark bits and synchronization codes were embedded inside selected frames in low-frequency DWT subbands using a watermarking scheme referred to as adaptive mean modulation (AMM).

In this study, to enhance the performance in robustness and processing efficiency, we follow the DWT-based framework developed in [19] and further introduce two efficient schemes for embedding synchronization codes and binary information in tandem. The remainder of this paper is organized as follows. Section 2 discusses the watermarking framework, whereby information bits and synchronization codes are embedded in selected DWT subbands. The way of implementing frame synchronization is also addressed here. Section 3 describes how to apply a spectrally shaped filter to achieve perceptual-based additive modulation for synchronization code embedding. Section 4 presents an improved AMM scheme for executing robust watermarking in the 2nd level approximation subband. Section 5 gives experiment results regarding the quality evaluation of watermarked speech and the watermark robustness against commonly encountered attacks. Conclusions are drawn in Section 6.