Brainstorm
October 13, 2009
Forrest Briggs
What can we do with 1 TB of audio data from H.J. Andrews?
Attending:
Faculty:
Bob Smythe
Fred Swanson
Tom Dietterich
John Selker
Julia Jones
Xiaoli Fern
Postdocs:
Rebecca Hutchinson
Students:
Uran Chu
Max Brugger
Sarah Frey
Arwen Lettkeman
Torrey Johnson
Jay Zarnetske
Terry Frueh
James Johnson
Ethan Derezynski
Tracy Kugler
Phoebe Zarnetske
Robin Paulekas
Paul Wilkins
Kendra Hatcher
Alan Tepley
Forrest’s presentation:
Songmeters – distribution in the HJ Andrews
Spectrograms
Noise reduction
Isolating bird song from other noises
Identification of bird species by sound
Active learning may help discriminate
Progress steady toward automated classification
CS and EE undergraduate capstone projects (CS students to build website to enlist the community for bird identification on these recordings, EE students to build better songmeter)
Suggestions for EE undergrad project: temperature sensitivity, battery life, standardize the orientation of the songmeter, numbers and directions pointing of the microphones, temperature sensor, humidity sensor, light sensor, fisheye photos of the canopy above. Expand frequency to detect bats? Program sensor to wake up and expand its frequency spectrum in response to sound? Solar power? Cost effectiveness? Will sounds be normalized?
Ethan: how to discriminate among individual birds? Forrest: separated in time or frequency
Tom Dietterich, Uran Chu: multiple microphones for detecting locations.
Fred Swanson: relationship of this work to monitoring seismic activity. Also, soundscape of streams and roads may be important to birds like spotted owl and breeding birds. Musicians inspired by sound.
John Selker: what are the technological limitations for this work? What would you spend additional money on? Forrest: not on computers. More microphones would be the best to spend it on, even though that would produce more data. More people would be helpful, other perspectives.
John Selker: Keck Foundation fiber acoustic listening device.
Forrest: biggest issue with current sampling is that it produces unprecedented temporal resolution, we are lacking spatial resolution.
Julia Jones: How large is area sampled by songmeter? Sarah Frey: less than 100 m radius sampled by people.
Phoebe Zarnetske: Are you matching the recordings to human inventories? Sarah, Forrest: yes.
Tom Dietterich: what affects the detectability of the species? Rain, wind, presence of people, traffic.
John Selker: there are devices that can measure rain acoustically that could be added to your measurement system. How will your work fit into the literature?
Forrest: major approaches to classification have focused on song level or syllable level organization of features. At syllable level, have used average spectra, other methods used for human voice recognition that alter the frequencies to better match how humans perceive pitch. Goal to develop features that are robust to various signal:noise ratios. At song level, Markov models subdivide song into syllables. Alternative is audio dictionary approach. Idea is that sounds to be classified may vary in length, using audio codebook approach subdivide the sound and create a histogram of “codes” that compose that sound.
John Selker: what about wavelets? What about using a classifying wavelet that looks like a bird call?
Fred Swanson: are you recording at frequencies beyond what humans can hear?
Forrest: In HJ Andrews songmeters are configured at 16 Khz, hear sounds up to 8 Khz. Probably don’t need more than 24 Khz to get birds.
Fred Swanson: could frequencies be adjusted to detect bats?
Phoebe: what about putting the microphones at different altitudes in the canopy?
Forrest: would be cool for fiber optics.
Sarah Frey: one of the questions I am investigating is how birds respond to structural diversity in the canopy.
Fred: Stuart Gage’s work at Wind River, Dave Shaw’s involvement.
Tom Dietterich: how to sample the 3D with a fiber optic cable?
Uran Chu: how does the signal:noise ratio vary with the frequency? Could sensors wake up and expand the frequency range in response to sound?
Forrest: the signal:noise ratio improves at higher frequencies. But the power of the higher harmonic is lower.
Ethan: Do you have an alignment problem? Forrest: arbitrarily I chose 10 seconds, this sometimes cuts off a syllable.
Tom: can’t you automatically segment out the syllables? A hierarchical approach might work better to eliminate empty chunks. Forrest: the segmentation is a challenge. One approach is to plot by energy before segmentation, but this is difficult with background noise.
Ethan: this is where the wavelet approach might help, by identifying the separation in time.
Rebecca: Are you using the time of the recording to help in the segmentation?
Forrest: Yes, that could be helpful. Most of the approaches are not supervised, but the example I showed is supervised, because I gave the algorithms particular regions to focus on and it was trained on other, known sounds.
Ethan: what caused the decline in classification from 6 to 20 species?
Forrest: 20 species of warblers that sound similar; 1/10 second recordings are tiny blips of sound.
Sarah: using the environmental variables to classify might be going backwards in terms of predicting where species are. Might not want to have too strong priors.
Tom Dietterich: What kind of species distribution model do we want here? We want to exploit the temporal resolution of the data.
Forrest: Questions we have already thought of are – when and where are birds arriving. This would be unprecedented.
Sarah: in some cases the recorders were deployed before arrival, but not in others. Another question is how the birds are moving through the landscape over the course of the seasons; this would improve over the 6 times of human sampling which is already high.
Tom: what would be a hypothesis?
Sarah: upper elevations begin by being snow covered and might expect birds to move into them
Ethan: what’s the density of bird song? Forrest: at 5 AM it is very dense, other times it is very quiet.
Max: what types of birds do you think you are seeing? Natal dispersers? Singing males?
Tom, Xiaoli: individual bird identifications.
Tracy: what about interspecies interactions? Are some species avoiding each other?
Tom: Could try an integrated model of all bird activity over the course of the day; predict when the sun hits a particular site.
Forrest: graph of number of calls by time of day that could be related to site characteristics.
Sarah: songmeter sites have temperature dataloggers.
Xiaoli Fern: one reviewer response from the pre-proposal was the short term research questions that can be answered. This discussion has raised some of these important short-term questions that could be used in the bioacoustics proposal.
Tracy: by short term what do you mean?
Xiaoli: questions that can be answered during the 4-5 year time frame of this work.
Tom, Sarah: competition for sound. How much spatial prospecting is going on?
Sarah: It’s not always known which birds are prospecting. Formerly it was thought that unpaired males would sing all season long, but in Adam’s and Matt’s experiments, the late-season male singing may be an indicator of breeding success.
Xaioli: so a site where lots of singing occurs in the late season should be one where birds are located next year.
Tom: the first sighting (of moths) seems to be a very noisy statistic. Some more robust measure such as a cumulative percentage may be more appropriate.
Tom: what about flight calls? Can birds be detected before they arrive?
Sarah: would depend on microphone arrays to detect flight calls.
Forrest: Should the song meters be left out during the winter?
Arwen: power suggestion: a battery on/off sensor? Forrest: songmeter sleeps in between scheduled recording sessions.
Max: solar power?
Tom: fiber optic cable most promising for power saving.
Alan:
Paul: will follow up with Forrest.
Kendra: geo 548 class will create a conceptual model of the acoustic landscape.
Forrest: do songmeters normalize the intensity of their sound? Does a songmeter 5 m from the stream produce a similar sound as one 10 m from the stream?
Robyn: first second of recordings appear to be louder than later.
Sarah: when loud bird sings, other quieter bird songs seem even quieter.