Difference between revisions of "Mass project"

From CCRMA Wiki
Jump to: navigation, search
(August 21, 2006)
Line 130: Line 130:
  
 
=== August 21, 2006 ===
 
=== August 21, 2006 ===
 +
Tuesday 9:30AM  '''Japan''' - Monday 5:30PM '''Stanford'''
 +
 +
 +
=== August 28, 2006 ===
 
Tuesday 9:30AM  '''Japan''' - Monday 5:30PM '''Stanford'''
 
Tuesday 9:30AM  '''Japan''' - Monday 5:30PM '''Stanford'''
  

Revision as of 18:08, 21 August 2006

Welcome to the Masking Ambient Speech Sounds project Wiki.

Experiments

Beta Test

The first listening tests will involve project staff members to check if things make sense. If it looks good we'll start working with non-project volunteers. Experiment 1, in the the CCRMA "Pit," will take about 30 mins. and involve 30 trials. There will be 6 conditions of masking sound crossed with 5 conditions of speech sounds. The masker (FM noise) and the speech sounds will be presented as if the sources are outside the room. We'll use the measured room model from Tokyo and the exterior sound source position (hallway). The "as if" impression will be created by convolving with the measured impulse responses.

Necessary ingredients: (x = done)

  1. (x) ambient room sound recording from Tokyo
  2. (x) 15 sec. recordings of FM noise masker with parameter variation
  3. (x) 4 min. recordings of 4 conversations (animated / not-animated, crowd / pair, always 50% gender balance)
  4. (x) 15 sec. clips cut from conversations
  5. (x) convolved versions of 15 sec. files putting them "as if" in the hallway
  6. (x) GUI for running randomized listening, A/B forced choice, logging results

Exp1GUI.png

Strategies to define conditions for FM masing noise

To define the conditions of this first experiment, the approach will be to leave all the parameters fixed, except the modulation frequency.

Noise set Contains a complete technical documentation of the masking noise generation. It also contains the soundfiles.

The conditions of the masking FM noise will be defined by the following criteria:

  • 3 bands of FM noise will be used (centered at 200 350 and 500 Hz):
    This bands are selected based on an analysis of speech voice recorded in the Tokyo office. The motivation behind this decision is to identify the relevant parameters in the leaking voice. For example, we know that the wall is filtering much of the high frequency components, so that's relevant in the selection of the main frequencies.
  • The amplitude (volume) of each band will be fixed:
    The amplitude was tuned in order to psychoacoustically balance the level of the three noise bands that will be used. This balance was done without modulation.
  • The amplitude of the modulation will be proportional to the modulation frequency:
    The motivation behind this choice is to minimize the annoyance effect. When the modulation rate is low, higher amplitudes are more noticed and annoying.
  • The relation between of modulation frequency of the 3 bands is then the main factor to define the conditions:
    For this experiment, 3 modulation rates are selected, 2, 5 and 7 Hz. The idea is to span some of the frequencies in the range of 2 to 7 Hz. Basically, all the combination of these 3 rates are used for each center frequency, plus a case with no modulation at all.


--Jcaceres 17:09, 24 July 2006 (PDT)

Beta test TODOs

The beta-test of the experiment tool took longer than anticipated. Some minor fixes remain. The ones I remember from yesterday (Friday, 28th) and the ToDo list for Monday (x = done):

  1. (x) delete input slider from bottom of GUI (in Qt Designer), final product should look like the picture above
  2. (x) when user hits "OK, Next" button, clear all the radiobuttons, with radiobutton->setDown(false)
    This worked out with setChecked(false) (inside a method, not in the connection)
  3. (x) comment out all the "cout" statements that are printing during trials, except for the one that says "behind"
  4. (x) find a sticky way to keep machine speed at max during trial (automatic energy saving may be the reason for the occasional stuttering)
    Jason comments:
    /usr/bin/cpufreq-selector -g performance
    you will select the "performance" governor and the cpu speed should go to the max and stay there.
    /usr/bin/cpufreq-selector -g userspace
    will return the governor to the original "userspace" governor.
    And:
    /usr/bin/cpufreq-selector -f 1000000
    will get the processor to the slow idle speed.
    From there the speed should again be "on demand". Regretfully it looks like sometimes the background daemon ("cpuspeed") gets fooled by these changes and dies. At least you can control all of this manually.
  5. (x) convert QString to const char for logger class file open (use const char * QString::latin1 ())
  6. (x) create a "shuffle" sort method in MainDialog.cpp and apply it for the actual first test
  7. I think each individual mono file repeating is ok, but I'm worried that they could slip out of sync. Don't know for sure. Better if the repeats for a group of four is from the first channel's repeat
  8. add envelopes at all file starts, stops, repeats (with STK's Asymp class), pipe the file's output through it
    I still need to add this, but I think after the first experiments, what we really need to do is make much longer files so they don't repeat, and the listener don't get a queu from that repetition.
  9. IF I've created a problem for disk files keeping up, you will see the message "behind" printed from FileWvIn and it will start stuttering, the next fix to try (and this might be important anyway for our sanity) is to go to quad files rather than 4 mono files for each layer.
    This doesn't look easy, I think I have to modify the entire Jukebox.cpp class in order to get this working...
  10. (x) Add a dialog in case the user doesn't select an option.
  11. (x) Change the silence always on A. Modify also the "correctness" of the selection, now is always set to be in A.
  12. (x) Turn off Sounds (alternative A and B) when user goes to next case.
  13. (x) Program is crashing at the end (it's quiting badly). If you go until the end, is not writing anything to the ouput files. If you stop it in the middle, it works. I may be probably a problem with some destructor...
    I get the problem with these test files:
    /usr/ccrma/snd/jcaceres/yamaha/recordings/experiments/experiment01/TEST/
    the message is:
    terminate called after throwing an instance of 'std::bad_alloc'
    what(): St9bad_alloc
    Aborted
    It works fine with these set of files:
    QString rootDir ("/usr/ccrma/snd/jcaceres/yamaha/recordings/experiments/experiment01/");
    FOUND IT!!! It was a problem reading a vector in MainDialog::setTrial (int n)


_____________ there are probably more things I'm forgetting, but this is close
GOOD LUCK!

--Cc 09:42, 29 July 2006 (PDT)

Findings on the Beta Test

  1. There is a low frequency of the voice that now is not beeing masked.
  2. We need to use a really long conversation, that does never repeat during the experiment.
  3. This corpus of conversations need to have "stationary" properties.


Bottom lines

  1. We're going to use just one room (Tokyo Office)
  2. We keep the 4CH setup.
  3. Spatialization ???

Conference Call Meetings

July 18, 2006

  • FM Modulation discussion (Yasushi's Comments, with Juan-Pablo's comment on answer A:):
  1. Do you have any idea how to specify frequency modulation for each frequency band?
    • A: based on speech freq, ~2-8 Hz
  2. The period in time for each frequency should be the same?
    • A: No, different. When it's the same the masking efficiency decreases. It seems also more anoying.
  3. Modulation speed will be getting faster according to higher frequency, or
    • A: I don't know yet, this is going to be the main parameter in the first experiment I think.
  4. The frequency modulation considering the voice sound
  5. We have to analyze how the voice sound is modulated in different frequency bands?
    • A: I thiks this is the best way, and we have to consider that the wall is filtering almost all the high frequencies.
  • Discussion of the experiment setup.
  • Look at the documentation, the new example of impulse responses, and delay of arrival.

July 24, 2006

Tuesday 9:30AM Japan - Monday 5:30PM Stanford

  • Discuss Experiment 1.
  • Ask Atsuko about calibration files and SPL meeter.
  • Comment diffusion in the Pit with PZM system (Hiroko).
  • Discuss Experiment Design writen by Hiroko and Atsuko.

July 31, 2006

Tuesday 9:30AM Japan - Monday 5:30PM Stanford

  • Discuss Experiment Design writen by Hiroko and Atsuko.
  • Explain experiment setup.
  • Discuss Atsuko's agenda at CCRMA.
  • Goals for this week are to finsih the setup (C++ and pit room) and collect and analyse some data in a couple of subjects.

August 21, 2006

Tuesday 9:30AM Japan - Monday 5:30PM Stanford


August 28, 2006

Tuesday 9:30AM Japan - Monday 5:30PM Stanford

Parameters for the Noise Generation

Variables

  • modulation width (critical band or speech sounds)
  • modulation rate (0.01 - 0.1 fc)
  • sinusoidal or stochastic modulation

Already fixed

  • with broadband noise (what shape, and how loud? - according to the speech)
  • band width of the noise (critical band)
  • amplitude of each channel (speech sounds spectral distribution)
  • number and frequency of center frequencies (3)


--Hiroko 18:27, 31 July 2006 (PDT)

Experiment design

what experiment to do

  • efficiency test
    • Stimuli: speech is mixed at randomized places in a stream of masking noise
    • Task: "hit the space key when you hear a speech"
    • Speech: 5 numbers (one, two, three, four, eight) spoken by a male and a female of different accents. Numbers were chosen so that they cover five vowels.
    • Masker: Genetic algorithm approach with human response. We vary one parameter first and then find one or two "sweet spots." Fix the parameter to those found values and vary the next parameter. Choose the best two - repeat this process.
    • Analysis: Response rate (response rate is low when speech is masked, we expect.) Response time distribution (more response time when speech is better masked, we expect.) Both analyses can be done within-subject and across-subject. We can also observe what kind of speech is better masked with a particular masking noise.
  • intelligibility test
    • Speech sounds: Idiomatic phrases and isolated words (Each masking sound has a phrase and word) - TBD
    • 4 sec per stimuli (15 stimuli/min)
    • Measure audibility and intelligibility thresholds
    • Better masking noise / parameter region are chosen from the efficiency test.
    • The answers the subject choose from are: "I don't hear speech" "Speech is audible" "Speech is intelligible"
    • complete random order (beyond the group of (1) sentense/words (2) masking noise types and (3) playback level)
    • we prohibit a sequencial presentation of a same stimuli from intelligible to less intelligible. (Same masking noise being loud to quiet)
    • analysis - we do not check the correctness - we only measure the intelligible impression
  • annoyance test
    • ten kinds of masking noise, silence, white noise with intruding noise, presented from 4 loudspeakers around.
    • Each one goes on for 30 seconds (or any length) fading in and out for 5 seconds.
    • Fade in the masking noise. Start with the word list, mental math, beep and repeat the word list. Fade out and fade in some enviromental noise (office, traffic, college cafeteria etc.), then next masking noise.
    • Word list is presented to the subject from a loudspeaker in front at 60 dBA.
    • Task: a word list is presented at the start. A subject does mental math for 30 seconds (6-10 questions.) After the beep, the subject has to recall the word list presented at the start. Masking noise switches with fade in/out with an environmental noise. Do the same task with the next masking noise.
  • Final comparison
    • For best 3 masking noises, mix in the typical conference noise (speech, paper shuffle, chair noise, typing sounds, and intruding noise) and ask the subjects which one sounds more "inviting."

(updated) Efficiency Experiment design

I've programmed up experiment 3. This uses the Santa Barbara corpus clips in a design that produces a percentage measure of masker effectiveness. It's for one masker (the best one arrived at from experiment 2) at a fixed playback level.

Jason has convolved the first SB dialog file, so it plays from the "hallway."

The subject hears a 2 second clip which the app selects randomly from the convolved file. As it's playing the app records the maximum RMS of the first channel of the clip. The subject responds with "yes" or "no" buttons according to whether they heard voices. The app records the response and the maximum RMS played, and then loops, playing the next randomly chosen 2 second clip.

This iterates a whole bunch of times over 5 minutes producing easily 50 trials per subject. The analysis plots the percentage of yes response vs. RMS. We should see a threshold RMS below which the clips were effectively masked.

For the final "efficiency rating" we go back into the convolved dialog file and calculate the percentage of time the signal is below the threshold.

Atsuko's visit Agenda

  • Friday August 4th,
    1pm Meeting (listening room)
    5:30pm - Conference Call Japan
  • Saturday August 5th
    Noise, narrowing parameters.
  • Sunday August 6th
    Meetings with Jonathan Berger and Hiroko
  • MondayAugust 7th
    Pscychoacoustic generic tests (Hiroko)
    Brain storm spatializtion parts - experiments strategies
    Meeting, Jonathan Berger, Jason, Juan pablo, Atsuko and Hiroko.
  • Wednesday 9th,
    1pm - Meeting
    5:30pm - Conference Call Japan

Links