MFCC-based perceptual hashing for compressed domain of speech content identification

Transcription

1 Available online Journal o Chemical and Pharmaceutical Research, 014, 6(7): Research Article ISSN : CODEN(USA) : JCPRC5 MFCC-based perceptual hashing or compressed domain o speech content identiication Qiu-yu Zhang 1 *, Yang-wei iu 1, Yan-jun Di 1, Qian-yunZhang and Peng-ei Xing 1 1 School o Computer and Communication, anzhou University o Technology, anzhou, China School o Communication & Inormation Engineering, Shanghai University, Shanghai, China ABSTRACT Current research on speech content identiication aim primarily at raw wideband speech signals, which are generally transmitted in a compressed ormat. This makes it unable to meet the demand o speech content identiication in compressed domain. This paper proposes a new speech perceptual hashing algorithm or speech content identiication with compressed domain based on MFCC (Mel Frequency Cepstral Coeicient), to solve problems o real-time speech content identiication and large quantity o voice message inormation over the mobile Internet. This algorithm extracts MFCC eature based on the raw wideband method. The process begins by extracting the MDCT coeicients, which are the intermediately decoded results o compressed speeches in MP3 ormat. These coeicients are translated to MFCC parameters and the binary hashing values are then generated rom these parameters, combined with human auditory eatures. This algorithm uses highly compressed data to realize ast identiication or speech content. Experimental results show that the proposed algorithm can realize tampering localization and increase 5% in eiciency when compared with raw wideband algorithms, with the precondition o robustness and discrimination. Key words: Speech content identiication; perceptual hashing; compressed domain; MFCC eatures; robustness INTRODUCTION With the development o inormation technology, the authenticity and integrity o voice products have been questioned when tools or digital media editing are processed [1]. Numbers o speech content identiication algorithms in the compressed domain are much less than traditional ones based on non-compressed ormat. A perceptual hashing is an easily computable unction that maps digital multimedia data into a compact digital digest. These unctions are widely applied in inormation security, where they are used as new algorithms or identiication, retrieval and identiication over an opening and unreliable network [, 3]. Since the parametric speech coding is completely dierent rom the way o audio compression, the audio hashing algorithm is unsuitable or speech algorithm [4]. Current researches on speech perceptual hashing usually take the original speech data as input. This large computational complexity can t meet the demand o real time application in speech communication terminals with limited resources [5]. In Re. [6] the author proposed a method with MEP (mixed excitation linear prediction) coding, using some parts o speech bit streams to generate hashing values. With malicious content-tampering discriminating abilities and less computational complexity, this algorithm is suitable or real-time system o speech content identiication. In Re. [7] proposed a perceptual hash algorithm designed or AAC (Advanced Audio Coding) audio to keep MDCT-based algorithms highly robust to compressed audio and provided a solution or speech content identiication in a compressed domain. Major process o traditional algorithm or the compressed speech data is as ollows [8]. The process begins by decoding the compressed speech into raw wideband data (PCM). Features rom decoded rames are then extracted 379

2 and urther analysis to achieve content identiication is inally made. Yet it continues to have a law o computational eiciency and complexity, thereore it can t aord real-time processing. Digital audio in practical applications is usually encoded in compressed ormats such as MP3, with the purpose o less data size, higher quality and easier transmission. Thereore research on compressed domain has positive signiicance [9]. For this reason we propose the ollowing algorithm to extract the MFCC eatures using human auditory system and MPEG audio coding. The process begins by decoding the MP3 stream and translating MDCT coeicients rom intermediate parameters. The MDCT coeicients o each rame are then translated to a 15-dimension MFCC coeicient vector ater Mel iltering. Content integrity certiication is inally veriied by matching the extracted hashing values. MDCT COEFFICIENTS IN COMPRESSED DOMAIN As a major vector eature o speech in the requency domain [10], MFCC is robust because o its ull consideration to the human auditory. In this paper we translate MFCC eature in compressed domain and select it as characteristic parameters. Physiological studies have shown that the human ear is very sensitive to the requency o audio, especially in the range o 00~5000 Hz [9]. A eature vector can be calculated by the original content o audio, but not MP3 compressed version o that content because o its process such as iltering and MDCT transormation. We extract the eatures rom MDCT coeicients which are intermediate parameters when decoding an MP3 ile. As a way o time-requency transormation, the MDCT-based method has been widely used in encoding audios, such as MP3, AAC, etc. [11]. In accordance to MPEG standards, audio stream is encoded rame-by-rame. A MP3 rame consists o granules and each granule contains 576 samples per granules [10]. We can get MDCT coeicients either by decoding each rame, or by perorming a modiied Discrete Fourier transorms on the 3 sub-band PCM (Pulse Code Modulation) signals. Each o the sub-bands corresponds to 18 MDCT coeicients. It has been proved that MDCT coeicients can be acquired through linear superposition o original signal (with weighting windows) and the aliasing signal perormed with SDFT (Shited Discrete Fourier Transorms) [1]. Moreover, with the assumption that there is no time shiting and the requency shiting is 50%, we can consider the original DFT as the nature o MDCT coeicients through linear transormation. Thus make it possible or us to extract perceptual eatures rom MDCT coeicient, or the reason that it is an approximate version o requency domain eature when we process audio stream using a sub-band ilter [13]. CONTENT IDENTIFICATION FOR COMPRESSED DOMAIN SPEECH A. Process o algorithm The MFCC-based perceptual hashing or speech content identiication is shown in Fig. 1.In this igure we get MDCT coeicients by processing the compressed speech with Human encoding. Fig. 1: Process o algorithm B. MFCC eature extraction MDCT coeicients can be regarded as an approximate version o linear spectrum o DFT [1]. Only considering the energy o these coeicients, we can extract eatures in the compressed domain according to MFCC algorithm in the non-compressed domain [14]. DFT coeicients with equal intervals are calculated ater MP3 hybrid iltering. The dierence is that none o these parameters divides the requency spectrum into orm o n. 380

3 We propose this eature extraction method on the basic that MDCT coeicient contains enough inormation to describe requency spectrum.firstly, we redeine the MDCT coeicient o each rame to six critical bands, which are similar with the critical brand according to its bandwidth. Then an n-dimension MFCC eature parameters is calculated via triangle iltering the MDCT coeicients. In this paper, the redeine o critical band is not taken into account o when extracting eatures. A ilter is perormed beore processing. Center requency o Mel iltering and ilter banks are determined based on the bits o calculated eature vectors. MDCT coeicient o each rame will inally be converted to a 15-dimension eature vector ater Mel triangle ilter banks ollowed by the cosine transormation. It has lower resolution when compared to 576 parameters DFT in original audio without compression, but can aord identiication or actual speech signals. The details o the proposed method are expressed as ollows. Step1:Intra-rame energy. Because o the noise and estimation error o spectrum, the logarithmic energy o MDCT spectrum is calculated ater a Mel ilter to improve robustness. Considering time-varying in actual speech signals, the calculation process as ollows is based on each rame. The square o MDCT coeicients in each two granules o one rame is now calculated. The corresponding energy is denoted by MDCT 1 and MDCT as shown in (1). MDCT = MDCT 1 + MDCT (1) The mean value is calculated and the energy vector with 576 elements is given, which is accord with equal interval distribution in requency domain. Step:Mel triangle iltering. Human perceptual auditory increases linearly with requency in the range rom 0Hz to 1000Hz, but they show a logarithmic relationship when requency is above 1000Hz. We deine 16 ilters corresponding to the centre o Mel requency to reduce computational complexity. The upper - lower limit o iltering requency (denoted by and H ) is mapped to the Mel requency and the range is determined in (). B( B( ) = 115 ln(1 + H ) = ln1 + H 700) 700) () In this ormula we deine a method to map the actual requency to the Mel requency, where B H represents the upper bound o Mel domain and B represents the lower. B Mel = B H B (3) Using (3) we arrive at a Mel central requency by dividing Mel bandwidth (B Mel ) into the number o ilters equally and mapping central requency o ilters to corresponding requency linear sequences. FC( m) = N F B 1 ( B( B( ) + m ) B( m + 1 H ) ),1 m 16 s (4) In (4) we deine B -1 (b) = 700(e b/115-1) as an inverse unction o B and modiied by inclusion o a actor o N/F s in order to map center requency to requency linear sequence. Here F s is sampling rate and N =576 equal the number o MDCT coeicients in each granule. The triangle ilter is a unction that calculates the component o requency domain in range o Mel requency and multiplies the MDCT energy amplitude by corresponding actors. Transer unction o Mel ilter is shown in (5). 381

4 H m k FC ( m 1), FC ( m 1) k FC ( m) k < FC ( m 1) or k > FC ( m) FC ( m) FC ( m 1) ( k ) = FC ( m + 1) k, FC ( m) k FC ( m + 1) FC ( m + 1) FC ( m) (5) The actors 1 (FC(m) - FC(m - 1)) and 1 (FC(m + 1) - FC(m)) can be seen as the iltering actors around center triangle iltering. These corresponding actors are dierent due to nonlinear bandwidth. Sequence number o ilters is denoted by m. See also the MDCT coeicient in Fig. 1, where k is corresponding to the coeicient ranging rom 0 to 575. Step 3:Energy ater iltering. The triangle ilter in the last step has a unction o requency division; thereore it can be used to process the energy coeicient in Step1. Given Noise Reduction, dynamic boundary o requency spectrum and distribution o logarithmic energy spectrum, the output o the ilter banks is calculated as (6). 575 X ( m) = ln( MDCT k = 0 H m ( k)),0 m 15 (6) Where m and k represent sequence number o ilters and MDCT (possibly are 0 over high requency). Step 4:Translation to cepstrum by DCT. In order to assure the ollowing decorrelation to dierent channels o MDCT spectrum, we perorm a DCT transorm on the output X(m) o ilters. Mel( n) 15 = m= 0 X ( m) cos [ πn( m + 0.5) 16] ),0 n 15 (7) A 15-dimension MDCT vector is acquired using (7). However it makes distinguishing dierence o these dimensions in content identiication. In this paper we select the whole vector except or the irst dimension considering its much less inormation. C. Hashing values extraction The 15-dimension coeicient vector is extracted rom single rame o the speech signal as described in section B. Because o the real-time demand o speech signal and computation complexity o extracting hashing values rame-by-rame, the 10-dimension vector o every 10 rames is divided into a sub-band. Only binary sequences translated rom eigenvector o these sun-bands are retained. This method in (8) keeps robustness and unidirectivity as well as reduces the data quantity. Formula (8) ormally deines the bits o the hashing string. The MDCT coeicient is denoted by Mel c (t, m), the m-th bit o the hash H in t sub-bands is denoted by H(t, m) and the threshold T is equal to zero. 1, Melc ( t, m) T H ( t, m) = 0, Melc ( t, m) < T (8) Finally we get a hashing block consisting o the m bit hashing string extracted rom 10 subsequent rames (6ms per rame) with the above algorithm. The minimum precision o identiication is 0.6s in speech content and tampering localization is achieved. D. Hashing matching Two derived threshold values denoted by τ 1 and τ (τ 1 <τ ) will determine whether two 3 second speech clips are similar or tampered, by compared to bit error rates (BER) o hashing values which are extracted rom the above clips. It will be declared either similar when BER is below a certain threshold τ 1, or tampered when BER is above τ. BER between τ 1 and τ calls or a tampering localization detection. 38

5 Qiu-yu Zhang et al J. Chem. Pharm. Res., 014, 6(7): In this paper, we present a ull procedure o perormance tests and their results. The database o speech clips in our experiment is shown in Table I. The experiments environment platorm is Windows7 operating system o Dell notebook, CPU is Inter Core i3-450m,.4ghz and G memories, MATAB R010b. A. Robustness analysing All o the 1000 MP3 speech clips speech signals. Resample consisting o subsequent down and up sampling to.05 khz and khz. Echo addition with attenuation o 60%. Increase the volume by 50%. Reduce the volume by 50%. ow-pass iltering using a ith order Butterworth ilter with cut-o requencies o khz. Thereater the hash values are extracted rom the speech clips which are processed with the irst ive content-preserving operations. The BER between the hash values perceptual content). RESUTS Table I: Speech Clips Sampling Rate Bit Depth Channel Bit Rate 44100Hz 16 bits mono 18kbps are processed as ollows. Each o them can preservee the perceptual content o is then determined. The resulting bit error rates are shown in Fig. (with same Fig.: BER in 500 clips with same perceptual content Here we arrive at a BER mostly below 0.3 rom clips with same content, which ensure the robustness o the proposed method. Robustness is related to the extracted perceptual eatures as well as the threshold value. Table II lists the ratio o clips declared equal using dierent threshold values. (These clips are subjected to dierent content-preserving operations). Table II: Passing Rate Threshold Volume down Volume up Echo Resample ow-pass Filter % 77.6% 75.3% 1.4% % 98.7% 99.8% 93.4% 97.9% 30.% 58.4% 85.4% 90.8% 383

6 Experimental results lead to the conclusion that we arrive at an extremely high identiication precision. It also keeps high robustness to operations o resample and volume reducing with a threshold τ o 0.3. B. Discrimination analysing In this paper we measure the discrimination ability or dierent speech contents with probability distribution because o the randomly variable BER. Fig. 3 illustrates the comparison o the distribution o BERs and the normal distribution. It shows that BERs has a normal distribution approximately. Fig. 3: Normal probabilityplot o BER among dierent speech content The two contradictory parameters FRR and FAR can be used to measure the robustness and o ability discrimination respectively in proposed algorithm. In dierent applications it poses dierent emphases and FAR has slightly higher priority in our scheme to discriminate dierent and tampered clips. Fig. 4 and Fig. 5 show the FRR-FAR curve o 500 speech clips which are randomly selected rom speech database. The cross point in Fig. 4 is cause by the weak robustness to low-pass iltering in the proposed method. The experimental results o other content-preserve processing are shown in Fig. 5. Fig. 4: FRR-FAR Curve with ow-pass iltering 384

7 Fig. 5: FRR-FAR Curve without ow-pass iltering It suggests that the proposed method is highly robust and able to discriminate between malicious content replacements and content-preserving operations, with the identiication threshold value τ o C. Perormance analysing The proposed method aims primarily at applications in communication terminals with limited resources. The eiciency o the algorithm can be measured with bit rate as (9) bps (9) In this paper 15-dimension hash string is extracted rom 10 rames (lasts 60ms) and leads to a low bit rate 115bps. This experiment process works on the platorm o MATAB 010b, using 100 speech clips. Each clip is encoded at a 18kbps bit rate and lasts 4s. Experimental results show that the eiciency is increased by 5% compared with other raw wideband algorithm, which aords real-time applications. D. Tamper location A 7s clip randomly selected and cropped with two clips larger than 10 rames. Experimental results o malicious tampering are shown in Fig. 6. Fig.6: Tampering Detection and ocation Human speech rate is about 15 words o one minute and 480ms each word. The hash string in the proposed algorithm is extracted rom 10 rames with time intervals o 60ms, which could be used to content tampering detecting and locating or one or more partial clips in speech signals. 385

8 CONCUSION In this paper we propose an identiication algorithm or integrity identiication o speech content in compressed-domain. This method is based on perceptual hashing algorithm and integrated with MFCC eatures, which are translated rom intermediate parameters when decoding, named MDCT coeicient. Hash values are extracted rom MFCC eatures based on raw wideband methods. The experimental results show that the eiciency is increased by 5% compared with other raw wideband algorithms. The robustness and ability o discrimination is also maintained. As the precision o 60ms, the proposed method could be used in real-time identiication as well as tampering detection and location. Based on the low cost o storage and computation we believe that this method has great value in certain applications. Acknowledgments This work is supported by the National Natural Science Foundation o China (No ), the Natural Science Foundation o Gansu Province o China (No. 11RJZA006, No. 1310RJYA004). The authors would like to thank the anonymous reviewers or their helpul comments and suggestions. REFERENCES [1] Y. H. Jiao andx. M. Niu. IEEE Signal Processing etters, 009, 16(9), [] X. M. Niu and Y. H. Jiao. Acta Electronica Sinica, 008, 36(7), (in Chinese). [3] GUPTA S, CHO S, KUO C-C J. IEEE Multimedia, 01, 19(1), [4] Y. H. Jiao. Research on Perceptual Audio Hashing[Ph.D. dissertation]. Harbin:Harbin institute o technology,010(in Chinese). [5] J. Gu. Research on Key Technologies o Speech Perceptual Identiication [Ph.D. dissertation]. Heei: University o Science and Technology o China, 009(in Chinese). [6] A. Shahbazi, A. H. Rezaie and R. Shahaazi. MEPe Coded Speech Hiding on Enhanced Full Rate Compressed Domain. In: 010 Fourth Asia International Conerence on Mathematical/Analytical Modelling and Computer Simulation, Kota Kinabalu, Malaysia, 010, [7] Y. H. Jiao, M. Y. i, B. Yang, et al. Compressed Domain Rubost Hashing or AAC Audio.In: IEEE International Conerence on Multimedia and Expo, Hannover, 008, [8] Gary Grutzek, Julian Strobl, Bernhard Mainka, et al. Perceptual Hashing or the Identiication o Telephone speech. In: Proceedings o Speech Communication, 10. ITG Symposium, Germany, 01, 1-4. [9] Y. H. Jiao, Q. i and X. M. Niu. Compressed Domain Perceptual Hashing or MEP Coded Speech. In: Intelligent Inormation Hiding and Multimedia Signal Processing, Harbin, China, 008, [10]. Y. Chang and X. Q. Yu. Journal o Computer Applications, 009, 9(4), (in Chinese). [11] Y. J. Wang,. Guo and C. P. Wang. Journal o Chinese Computer Systems, 011, 3(7), (in Chinese). [1] Y. Wang, eonid P. Yaroslavsky, M. Vilermo. On the Relationship Between MDCT, SDFT and DFT. In: Proceedings o the 5th International Conerence on Signal Processing, Beijing, China, 000, [13] Y. iang, C. C. Bao. Acta Electronica Sinica, 01, 40(6), (in Chinese). 386