THE USE OF A FORMANT DIAGRAM IN AUDIOVISUAL SPEECH ACTIVITY DETECTION K.C. van Bree, H.J.W. Belt Video Processing Systems Group, Philips Research, Eindhoven, Netherlands Karl.van.Bree@philips.com, Harm.Belt@philips.com ABSTRACT We present an audiovisual approach to the problem of voice activity detection for systems with a single microphone and a single camera with multiple people in the camera s field of view. We aim to have a speech activity detection result per person. The approach utilizes a face tracking and lip contour tracking algorithm for the video analysis, and pitch presence detection and formant frequency tracking algorithms for the audio analysis. When from the audio we detect speech activity and from the video we find lip activity for more than a single person, we check for each person whether the vowels correspond with the video mouth parameters to find out if this person speaks. To this end we make use of the F 1 - F 2 speech formant diagram in which we propose three vowel groups that are distinctive both from audio and video data. 1. INTRODUCTION For many speech signal processing applications such as speech telecommunications and speech recognition systems it is relevant to be able to detect speech activity. Speech activity detection algorithms like in [1] work well under good acoustic conditions but suffer from false detections when ambient noises are speech-like. Activity detection techniques purely on video lip motion like in [2, 3] aim to be independent of these noises, but suffer from false detections when people move their lips in facial expressions without talking. In this paper we adopt an audiovisual approach to the task of speech activity detection. We consider the case of multiple persons in a camera s field of view with only one of them talking at a time, while others could be moving their lips without talking. We aim to have a speech activity detection result per person. As summarized in Fig. 1, we propose to correlate speech features with mouth features to find proof of which person utters a detected vowel. In Section 2 we focus on the audio modality. An audio speech activity detector is given, and a 2-dimensional formant diagram is introduced in which we propose to distinguish three well-separated vowel groups. Section 3 deals The authors thank the reviewers A.C. den Brinker and R. Jasinschi from Philips Research Laboratories in Eindhoven, and R. Sluijter from the Eindhoven University of Technology for their useful comments. Figure 1. Flow chart of the audiovisual speech activity detector with the video modality. We present a lip detection and tracking algorithm. We introduce in the audio formant diagram distinguishable mouth shapes with the three vowel groups that we selected and use this in a lip detection algorithm. The main contribution of this paper is the specific choice of this diagram, and its application for the improved audiovisual activity detector presented in Section 4. Finally, in Section 5 we give our conclusions. 2. AUDIO VOICE SIGNALS AND DETECTION 2.1. Audio Speech Activity Detection We next describe the applied audio-only speech activity detector. In the first step we divide the signal into frames by windowing. Next, for each frame we investigate whether there is non-stationary signal activity. If so, the final step is to verify the presence of pitch. As such we will get one detection per audio frame. Let s[n] denote the sampled audio signal and B the audio frame size. We take B = 128 at F s = 8 khz. Let S w [k] be the M-points discrete Fourier transform (DFT) result of the Hanning windowed 2B last audio samples. We take M = 2B. P s [k] is the power spectrum: P s [k] = S w [k] 2. Note that, due to the symmetry in the frequency domain, only the first M/2 + 1 points of P s [k] are relevant. We estimate 2007 EURASIP 2390
from P s [k] the stationary background noise part P n [k] with a minimum statistics method [4]. We detect non-stationary signal activity when the SNR exceeds θ (we take θ = 8): M/2 k=0 P s[k] P n [k] M/2 k=0 P n[k] > θ. (1) The auto-correlation ρ[l] is calculated by the inverse DFT of P s [k] P n [k]. L = {l min,, l max } is the lag range corresponding to the frequency range of human pitch (between 80 and 500 Hz). Like in [1] we assume presence of pitch when ρ[l]/ρ[0] > θ ρ for any l L. (2) A good value for θ ρ is 0.75. We detect speech when signal activity is detected according to Eq. 1 and when pitch presence is detected according to Eq. 2. To deal with consonants we keep the detection result positive for a small extended time period when pitch is no longer present and Eq. 1 is still satisfied. The audio-only speech detector works well for one person but it cannot discriminate between different people. 2.2. Voice Signals and Vowel Groups Next we link speech formant frequencies to vowels. The formant frequencies are denoted by F 1, F 2,.... In [5] (Figure 9), Peterson and Barney plot for ten vowels uttered by 76 speakers the location in an F 1 -F 2 diagram, and they then distinguish ten smoothly-shaped regions for each vowel. The figure demonstrates that already with the first two formant frequencies one can reasonably well predict the uttered vowel. In our further discussion we therefore restrict ourselves to F 1 and F 2. To estimate F 1 and F 2, we first perform DC-removal and pre-emphasis filtering. The signal is then Hanning-windowed. For each windowed audio frame a 10-th order auto-regressive (AR) model is calculated [6]. To find F 1 we search for the first (lowest) frequency in the range of 200 to 830 Hz at which the AR spectrum peaks with a sufficiently high Q- factor. We do the same for F 2 in the range of 500 to 2650 Hz. Compared to Peterson and Barney, we confine ourselves to only three smoothly shaped vowel regions in the F 1 -F 2 diagram, see Fig. 2. We choose these regions to be wellseparated. By this we only consider vowels that are very distinct. The O-group contains vowels like /o/ and /u/, the A-group vowels like /a/ and /æ/, and the I-group vowels near /i/. Our specific choice for these three regions was based on our intuition, but is justified also by the results in (Table 1) of [7]. In this paper the authors have performed a neural network classification of vowels from reflection coefficients. Their results show classification confusions between vowels, reducing their classification accuracy to only three groups. These three groups are similar to our groups. In the next section we link the three vowel groups to distinct video lip shape parameters to improve our detector. Figure 2. Speech vowels in the F 1 -F 2 plane 3. VIDEO LIP SIGNALS AND DETECTION 3.1. Lip Finding and Contour Tracking We want to extract the vertical mouth opening m and the distance w between mouth corners as indicated in the lip contour model of Fig. 3a. First, we locate the faces in an image with a face detection algorithm based on [8]. Then, for each face, we select the mouth region-of-interest (MROI) as the lower part of the face region. An example MROI is shown in Fig. 3b. Figure 3. (a) Lip contour model; (b) Mouth region of interest (MROI) and search lines for lip edges. The locations of the mouth corners are extracted as follows. First, a binary image is calculated by dynamic thresholding of the MROI. We then look for the blob in the binary image which has the most mouth-like shape. Finally, the locations of the mouth corners are found as the left and right extremities of the mouth blob. The edges of the lips are found on search lines perpendicular to the line between the mouth corners (Fig. 3b). On the q-th search line we apply a function to the p-th pixel value such that a number R is obtained which is large for the red lip area and small for skin, teeth and the inner mouth. Then, for each pixel on the q-th line the value of R is compared to a threshold in order to yield the four edges of the two lips it crosses. Finally, two internal lip edges defining m are found as a second-order polynomial fit on the mouth corners and lip edge points, excluding outliers by a median operation. 2007 EURASIP 2391
We calculate R as R(q, p) = Q(q, p) {max(y (q, p) φ hi (q), 0) + max( Y (q, p) + φ lo (q), 0)}. (3) Here the number Q(q, p) is a mapping of the chroma values C b and C r according to Q(q, p) = α 1 Cr(q, p) + α 2 Cb(q, p), where α 1 and α 2 are chosen to favor the reddish color of the lips. We used α 1 = 0.88 and α 2 = 0.48. The second term in Eq. 3 is the correction for luminance Y on Q resulting in R becoming small for pixels that belong to the (bright) teeth or the (dark) mouth opening. The threshold φ hi (q) is calculated as the average luminance µ Y (q) of the q-th line. The threshold φ lo (q) is chosen to be selective only for the darkest pixels, and is calculated as φ lo (q) = µ Y (q) 0.8σ Y (q) where σ Y (q) is the standard deviation of Y (q). 3.2. Video Speech Activity Detection When speech activity is detected from the audio modality, we exclude activity of some people in the image by inspecting their video lip activity. The video detector must be conservative for silence detection, meaning when silence is detected it is quite certain that this is true. For visual speech detection we follow an approach like in [3]. Let t denote the index of a video frame. We detect speech activity from the video when ṽ[t] > θ v (4) and speech silence otherwise, with θ v a small fixed positive threshold that we obtained experimentally. Here ṽ[t] is the time-smoothed version of the vertical mouth velocity v[t] according to with ṽ[t] = α ṽ[t 1] + (1 α)v[t] (5) v[t] = m n [t] m n [t 1], m n [t] = m[t]/µ w, and µ w the average horizontal mouth opening in number of pixels serving the purpose of normalizing m[t]. Unlike [3] we use an asymmetric recursion in Eq. 5 with fast (rise) response favoring situations when the mouth is opening and slow (decay) response when the mouth is closing. In this way the detector is conservative for detection of silence. We also apply a fast response when the vertical mouth opening is completely zero. To achieve this α is given by { αf when m α = n [t] > m n [t 1] or m n [t] = 0, α s when m n [t] m n [t 1] and m n [t] 0. Here α f = τ f F v 1 τ f F v and α s = τ sf v 1 τ s F v, with F v the video frame rate and we choose τ f = 1/16s and τ s = 1/8s. A combined audiovisual speech detector for each person in the image is now obtained by multiplying the audio-only detection result from Section 2.1 with the video-only detection result belonging to that person from Eq. 4. With the addition of the video-only detector some people can be excluded when the audio modality has detected speech activity. It is not a sufficient measure however; there remains ambiguity when people move their lips without actually speaking (e.g. when they smile). We will show in the next section that some ambiguous detections can be eliminated by correlating detected audio formant frequencies with video lip shape parameters, which is the main contribution of this paper. 4. AUDIOVISUAL DETECTION 4.1. Lips and Vowel Groups In the F 1 -F 2 diagram of Fig. 2 we distinguished three vowel groups. Next, we relate these three vowel groups to typical mouth shapes. The vowel height is a feature expressing the vertical position of the tongue relative to the roof of the mouth during vowels sounds. Likewise, the vowel backness expresses the horizontal tongue position relative to the back of the mouth. In [5] vowels are related to F 1 and F 2, and in the International Phonetic Alphabet (IPA) chart vowels are related to vowel height and backness. More specifically, it can be deduced from [5] and the IPA chart that the first formant frequency F 1 is related to vowel height, and the second formant frequency F 2 is related to vowel backness. A low vowel from the A-group has a high F 1, and a high vowel from the O- group or the I-group has a low F 2. Back vowels from the O-group have a low F 2, and front vowels from the I-group have a high F 2. From video we cannot measure tongue positions, only lip shapes, but from the literature and our own experience we learned that there is a correlation between vowel backness (hence F 2 ) and roundedness of a lip shape. Also, we noticed from experiments a phonetic correlation between vowel height (hence F 1 ) and vertical mouth opening. The mentioned experiments involved the visual inspection of recorded lip images of persons which were pronouncing different isolated vowels. We selected from these experiments a representative lip shape in each of the three audio vowel groups (Fig. 4). 4.2. Detection Fig. 5 shows for two alternately-talking people the results of the audiovisual activity detector from Section 3.2, which is achieved by multiplying the result of the audio-only detector with the video-only mouth activity detector. The figure shows ambiguous detections. For example, in the interval t {19.0,, 21.0} only person 1 was talking, but the detector incorrectly finds speech for person 2 that momentarily 2007 EURASIP 2392
Figure 5. Audiovisual speech detection for two persons Figure 6. Audiovisual vowel detection moved the lips without producing sound. Using Fig. 4 we can remove some ambiguity when we find a clear visual support from the lip shape of only one person (and no other) for the detected formant frequencies. In this article we focus on the visual detection of roundedness because it proved to be the strongest cue. Roundedness is detected when m w > θ r and w < θ w µ w, (6) where θ r = 0.2 and θ w = 0.8 proved to give conservative results. When we detect from F 1 and F 2 that the current sound stems from the /o/-group, and when we detect roundedness for only one person according to Eq. 6, then we set the activity detection result to false for the other persons for as long as there is ambiguity. In Fig. 6 we plot the video lip parameters for both persons, and the detected vowels from F 1 and F 2. As shown in the vowel plot, at t = 19.6 a clear vowel from the /o/-group is recognized. As can be derived from the two roundedness plots and from Eq. 6 this vowel is visually supported by the mouth shape of person 1 and not by the mouth shape of person 2. 2007 EURASIP 2393
Figure 7. Improved audiovisual speech detection Figure 4. Distinct mouth shapes in vowel groups [3] D. Sodoyer, B. Rivet, L Girin, J.-L. Schwartz, and C. Jutten, An analysis of visual speech information applied to voice activity detection, in Proceedings ICASSP. IEEE, 2006, vol. I, pp. 601 604. [4] R. Martin, Noise power spectral density estimation based on optimal smoothing and minimum statistics, IEEE Trans. Speech Audio Processing, vol. 9(5), pp. 504 512, July 2001. [5] G.E. Peterson and H.L. Barney, Control methods used in a study of the vowels, Journal of the Acoustical Society of America, vol. 24(2), pp. 175 184, Mar. 1952. [6] S. Kay, Modern Spectral Estimation, Prentice-Hall, 1988. [7] S. Kshirsagar and M. Magnenat-Thalmann, Lip synchronization using linear predictive analysis, in IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2000, vol. 2, pp. 1077 1080. [8] P. Viola and M.J. Jones, Robust real-time face detection, International Journal of Computer Vision, vol. 57, no. 2, pp. 137 154, 2004. From this observation we can remove ambiguity by setting the detection result for person 2 to false immediately after t = 19.6 until the moment that there is no longer ambiguity. The resulting improved detection for the second person is shown in Fig. 7. 5. CONCLUSIONS We have given an audiovisual approach to speech activity detection for systems with one microphone and one camera, and with multiple persons in the camera s field of view. From an audio-only detector it is not clear which person talks. Combination with a video lip activity detector helps, but still leaves ambiguity when someone moves the lips without talking. We introduced a formant diagram in which we distinguished three separated vowel groups that can be linked with video lip shape parameters. We showed that this diagram is a useful tool to remove ambiguous detections and provide more clarity about which person talks. 6. REFERENCES [1] P.L. Chu, Voice-Activated AGC for Teleconferencing, in Proceedings ICASSP. IEEE, 1996, pp. 929 932. [2] P. Liu and Z. Want, Voice Activity Detection Using Visual Information, in Proceedings ICASSP. IEEE, 2004, vol. I, pp. 609 613. 2007 EURASIP 2394