Presentation is loading. Please wait.

Presentation is loading. Please wait.

Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia 2004 1 Radiation Directivity of Human and Artificial.

Similar presentations


Presentation on theme: "Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia 2004 1 Radiation Directivity of Human and Artificial."— Presentation transcript:

1 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Radiation Directivity of Human and Artificial Speech Teemu Halkosaari 1 & Markus Vaalgamaa 2 1 Helsinki University of Technology, Laboratory of Acoustics and Audio Signal Processing 2 Nokia Technology, Audio Quality Workshop of Wideband Speech Quality in Terminals and Networks: Assessment and Prediction, 8 th -9 th June 2004, Mainz, Germany

2 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Agenda 1.Motivation and Background 2.Goal 3.Measurements 4.Analyzes 5.Modelling 6.Results 7.Improvements to telephonometry 8.Conclusions Part I: WHY? Part II: The STUDY Part III: The OUTCOME

3 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Part I WHY? 1.Motivation and Background 2.Goal

4 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Motivation and Background Artificial mouths and HATS were originally designed for narrowband speech measurements with big landline handsets. In past years, the size of phone has become smaller and use of headsets is becoming common. In future, wideband phones are coming… It seems that HATS and artificial mouths will be used in future, although we do not know exactly the directivities of artificial mouths or human mouth especially on near field locations Additional interesting topic is that how does the speech content affect to the directivity? Extensive literature on the subject is not available Standardization is inadequate ITU-T P.58 Head and torso simulator for telephonometry ITU-T P.51 Arficial mouth narrow band: Hz, wide band: Hz Part I: WHY?

5 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Goal Measure frequency responses for artificial mouths to positions where microphones of handsets and headsets are situated. Repeat the same measurements for a group of test subjects Most important part of the study is to analyze the difference between artificial mouths and an average person. In practice a set of transfer functions is acquired. Part I: WHY?

6 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Part II The STUDY 3.Measurements 4.Analyzes 5.Modelling 6.Results

7 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Two separate systems Test subjects were measured in Helsinki University of Technology, anechoic chamber, 8 channel parallel recording system. Artificial mouths were measured in Nokia facilities in Salo, Finland, anechoic chamber, Audio Precision APwin measurement system, 2 channel parallel MLS measurement. Probes B&K 1/8 and ¼ inch measurement microphones 8 Sennheiser KE electret microphones In addition, some phones: large and small handset, and accessories: boom headset, mono and stereo headsets, were measured 3. Measurements: Arrangements Part II: The STUDY

8 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Measurements: Positions Referred to lip plane and its perpendicular on chest: axes from throat downwards and sidewards Two reference positions for transfer functions MRP = 25 mm in front of mouth for artificial mouths 0.5 m in front of mouth for test subjects. Measurement positions: Part II: The STUDY Near cheekIn front of mouth On chest

9 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Measurements: Setup for artificial mouths Part II: The STUDY Two B&K measurement microphones. Two impulse responses parallel (reference and point of interest) Apwin measurement system, MLS

10 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Measurements: Positioners Helmet and chest positioner were developed to keep microphones in right positions. Part II: The STUDY

11 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Measurements: Setup for test subjects (1) IOtech WaveBook/516 LapTop computer pre-amp E.A.A. 2 channels pre-amp E.A.A. 2 channels pre-amp E.A.A. 2 channels E.A.A. pre-amp channels Biasing unit channels Mixer DisplayVHS Recorder Camera Loudspeaker channels Test subject or HATS A/V-monitoring and Test Subject guidance Data acquisition system Microphones Monitoring roomAnechoic chamber LPT Part II: The STUDY Parallel 8 channel audio recording at 32kHz sample rate

12 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Measurements : Setup for test subjects (2) Part II: The STUDY Small anechoic chamber at Helsinki University of Technology and monitoring system

13 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Measurements: Test subjects Test subjects: 8 male, 5 female, mainly young students, Finnish as mother tongue Speech material: ”Kaksi vuotta sitten kävimme ravintola Gabrielissa Helsingissä ja söimme siellä padallisen fasaania banaanilla höystettynä” (includes all Finnish phonemes) Translation: “Two years ago we went to restaurant Gabriel and we ate there a pot of pheasant larded with banana “ Separate phonemes \n m  s r\ and \a e i o u y ä ö\ Different speech volumes: loud, normal, and silent Part II: The STUDY

14 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Analyzes Starting point: parallel 8-channel 32kHz 16bit raw audio data. Impulse responses in measurements for artificial mouths -> directly converted to transfer functions. Phonemes and active speech were segmented manually (CoolEdit Pro, around manual marks by the way ). All analyzes were made in Matlab environment: SNR and Coherences to ensure reliability Power spectra estimates Transfer function estimates Confidential intervals consideration for transfer functions FFT 1024 samples that corresponds to 32 ms, Hanning-window, 50% overlap Averaging and Smoothed to 1/3-octave bands Speech was averaged weighting with coherence Part II: The STUDY

15 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Modelling In addition to measurements two models were built based on acoustic field theory Results were validated by modelling: Sphere and piston source Head = sphere Mouth = radially vibrating round surface of the sphere (piston). An infinite baffle can be added to simulate the upper body. Mirror source method Numerical implementation in Matlab Part II: The STUDY

16 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Results: HATS Nearfield Measured and modelled results for B&K 4128 HATS Transfer functions from Mouth Reference Position (MRP) to other positions Part II: The STUDY Transfer functions from MRP to near cheek positions: Large handset (blue circles) Small handset (red squares) Boom headset (green triangles) Corresponding models (black dashed) Transfer functions from MRP to chest positions: Near throat with vest (blue circles, solid) Near throat without (blue circles, dashed) Middle of chest with vest (red squares, solid) Middle of chest without vest (red squares, dashed) Corresponding models (black dashed)

17 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Results: HATS Far field (1) Measured results with B&K 4128 HATS Transfer function from MRP to positions in the front of mouth: 10 cm (blue circles) 25 cm (red squares) 50 cm (green triangles) 100 cm (violet quadrangles) with (solid) and without vest (dotted) Part II: The STUDY

18 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Results: HATS Far field (2) Measured and modelled for B&K 4128 HATS Transfer function for positions in front of mouth on 50 cm distance. Vest on (blue circles) Vest off (red squares) Bare head, torso unattached (green triangles) Model with the baffle (black dotted, thin) Model without the baffle (black dasked, thick) Part II: The STUDY

19 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Results: Human – B&K HATS Comparison between transfer functions, referred to 0.5m in front of mouth position. Positions, colors and symbols are same as in the slide 16. TF Human / TF HATS 1/3 mouth cross-section area ratio in models 0 dB values indicate that there is no difference between human and HATS Positive values indicate that HATS is more directive, i.e. has less SPL on measured microphone positions on next to cheek or on chest. Next to cheekOn chest Part II: The STUDY

20 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Results: B&K 4227 and Head Acoustics HMS II.3 HATS Comparison between transfer functions, referred to 0.5m position TF Human / TF HATS Positive values indicate that HATS is more directive, i.e. has less SPL on measured microphone positions on side. Transfer functions from MRP to near cheek positions: Large handset (blue circles), Small handset (red squares), and Boom headset (green triangles) Next to cheek for B&K 4227Next to cheek for HA HATS Part II: The STUDY

21 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Results: Speech content Speech content changes the directivity pattern Speech style: loud – silent etc.. Articulation: continuous phonation – natural speech Phonemes (Smoothed Human/HATS difference for mic. pos. 1.3 below with 95% confidence intervals) Mouth aperture size = steepness of the response on high frequencies Statistical approach: ANOVA model did not apply = speech data scattered Part II: The STUDY TFs human vs. HATS by vowel groups

22 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Results: Reliability Beforehand by careful design of the measurement processes Reliability of human measurements: SNR and Coherence (estimates) SNR on wide-band: >25dB Coherenceon narrow-band: >0.8, on wide-band: >0.65 Normal distribution hypothesis 95% confidential intervals on wide-band: <±0.5dB Reliability of HATS measurements is better than for humans i.e. very good (and thus omitted here) Part II: The STUDY

23 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Part III The OUTCOME 7.Improvements to telephonometry 8.Conclusions

24 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Improvement proposal: Equalization (1) Artificial mouths (especially B&K 4128) does not correspond to average test subject Difference curves are ”well-behaving” for B&K HATS -> A low order filter could equalize the difference near cheek. Example: Yulewalk recursive IIR filter design (least-squares method) was applied in Matlab for smoothed and averaged differences (B&K HATS) Average of curves on right (blue) IIR N = 3 (red) IIR N = 7 (green) Part III: The OUTCOME Difference and 95% confidence limits between human and B&K HATS near cheek positions: Large handset (blue circles), Small handset (red squares), and Boom headset (green triangles)

25 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Improvement proposal: Equalization (2) Smoothed difference curves and 95% confidence limits between human - B&K 4227 (left) and Human - Head Acoustics HMS II.3 (right) on near cheek positions Large handset (blue circles) Small handset (red squares) Boom headset (green triangles) Part III: The OUTCOME Human - B&K 4227Human - HA HATS

26 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Improvement proposal: Mouth size B&K 4128 HATS Positions 1.3 (left) and 1.1 (right) where measured with 3 different mouth sizes small, 15 mm x 11 mm, adaptor partly blocked (blue circles) normal, 30 mm x 11 mm, adaptor (red squares) big, 42 mm x 16 mm, mouth adaptor removed (green triangles) Part III: The OUTCOME Position 1.3 Human - B&K HATSPosition 1.1 Human – B&K HATS

27 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Improvement proposal: Vest About 2cm thick vest should be used in B&K HATS measurements. How does it change the directivity in position on the chest? Let’s take off the vest and compare to average test subject (dashed lines) The near-chest reflection shifts to higher frequency Difference between human and HATS measurements on chest positions: Near throat with (blue circles, solid) and without vest (blue circles, dashed) Middle of chest with (red squares, solid) and without vest (red squares, dashed) Part III: The OUTCOME vest on (solid) vest off (dashed)

28 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Conclusions Key results: On high frequencies B&K HATS is too directional. Near cheek: the difference (Human/Artificial mouth) is slighly increasing by moving away from the lip plane. On averate the difference between human and artificial mouth is between 1 and 6 dB. Near chest: the differences in chest reflections are significant. Speech content affects the directivity. Reasons: Mouth cross-section area during speech is smaller than in artificial mouths. Speech content and mouth cross-section area are linked. The torso and head design of the artificial mouths do not correspond to the acoustical characteristics of a real body. What could be improved? Near cheek: simple equalization was proposed. Improvement of the structure of HATS (mouth size, torso, etc.). This path requires more measurements and analysis. Vest usage should always be considered carefully. Part III: The OUTCOME

29 Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia Agenda 1.Motivation and Background 2.Goals 3.Measurements 4.Analyzes 5.Modelling 6.Results 7.Improvements to telephonometry 8.Conclusions Part I: WHY? Part II: The STUDY Part III: The OUTCOME


Download ppt "Teemu Halkosaari & Markus Vaalgamaa: Radiation Directivity of Human and Artificial Speech 08-June-04© Nokia 2004 1 Radiation Directivity of Human and Artificial."

Similar presentations


Ads by Google