lion-city-soundscapes

This repository stores code to download the files in the Lion City Soundscapes (LCS) dataset available at https://doi.org/10.21979/N9/AVHSBX, as well as replication code for the following publication related to the dataset:

Kenneth Ooi, Jessie Goh, Hao-Weng Lin, Zhen-Ting Ong, Trevor Wong, Karn Watcharasupat, Bhan Lam, Woon-Seng Gan, Lion City Soundscapes: Modified Partitioning Around Medoids for a Perceptually Diverse Dataset of Singaporean Soundscapes, JASA Express Letters, 2024.

The LCS dataset contains audio-visual recordings made at the 62 locations identified by the Singapore Soundscape Site Selection Survey, where each soundscape is given an ID in the form S####, where #### ranges from 0001 to 0062. The set of audio-visual recordings for each location comprises the following, as both "full-length" recordings (ranging from 9-38 minutes in length) and 1-minute excerpts of the "full-length" recordings:

• 19-channel WAV file recorded using a ZYLIA ZM-1 3rd-order Ambisonic microphone
• 2-channel stereo WAV file generated from the 19-channel WAV file upon recording using a ZYLIA ZR-1 Portable Recorder
• 3-channel HDF file recorded using a B&K Type 4101-B binaural microphone (first two channels) and GRAS 146AE Free-field Microphone (last channel) connected to a HEAD Acoustics SQobold Data Acquisition System
• 2-channel WAV file recorded using the binaural microphone, as generated from the first two channels of the HDF file
• 1-channel WAV file recorded using the GRAS 146AE Free-field Microphone, as generated from the last channel of the HDF file
• 360-degree video file (in spherical format) recorded using an Insta360 One R camera and onboard microphone

Further information on the audio and video files in the dataset itself is available at https://doi.org/10.21979/N9/AVHSBX. For the visualisation code, please refer to https://github.com/ntudsp/lion-city-soundscapes-visualisation

Getting started

Firstly, clone this repository by manually downloading it from https://github.com/ntudsp/lion-city-soundscapes, or enter the following line from a terminal (you need to have git installed first, of course):

git clone https://github.com/ntudsp/lion-city-soundscapes

You may then navigate to the downloaded folder with

cd lion-city-soundscapes

If you are using conda as your package manager, you may enter the following line into a terminal to install the required packages into a conda environment (or you may install them manually using the requirements stated in lcs.yml):

conda env create -f lcs.yml

Activate the conda environment by entering the following line into a terminal:

conda activate lcs

(If you are running the code on a computer with macOS installed, and the above commands fail, try conda env create -f lcs-mac.yml and conda activate lcs-mac instead.)

To run the replication code and baseline models, as reported in our publication, you may enter the following line into a terminal (this opens a Jupyter Notebook in your default browser):

jupyter lab --notebook-dir . code/replication_code.ipynb

In addition, if you want to download the audio and video files making up the dataset (which includes the 1-min excerpts of all raw audio and video recorded for all soundscapes, as well as the full-length recordings for S0001, S0002, and S0041), you may alternatively enter the following into a terminal (this will download ~210 GB of data from the Internet):

cd ./code
python download.py manifest.csv

If all files have downloaded successfully, your directory structure should match this. The replication code does not require a prior download of any of the audio or video files from the LCS dataset to work.

Directory structure

This repository

.
├── code                          # Code used to process the raw data and output the results.
│   ├── download.py              
│   ├── download_utils.py        
│   ├── lcs_points.py            
│   ├── lcs_utils.py             
│   ├── manifest.csv             
│   └── replication_code.ipynb   
│                                
├── figures                       # Folder containing all reference figures used in the Jupyter Notebook (29 files; all png)
│   ├── all_soundscapes.png
│   ├── ...     
│   └── uniform_circle_points.png 
│
├── results                       # Folder to store reference results & results generated by the Jupyter Notebook
│   └── seed-12-epoch-79-ref.csv
│                                
├── CITATION.cff                  # Citation information for the dataset in plain text.
├── README.md                     # This file.
├── lcs.yml                       # The Anaconda environment containing required packages to run all the code in ./code (for Windows and Ubuntu)
├── lcs-mac.yml                   # The Anaconda environment containing required packages to run all the code in ./code (for macOS)
├── metadata.csv
└── predictions.csv

After running python download.py manifest.csv

.
├── code
│   ├── download.py
│   ├── download_utils.py
│   ├── lcs_points.py
│   ├── lcs_utils.py
│   ├── manifest.csv
│   └── replication_code.ipynb
│
├── data
│   ├── S0001
│   │   ├── S0001_1min_Insta360.mp4          # 4096 x 2048 resolution, min. 40 mbps @ 29.97 fps, audio min. 128 kbps, 2 channels, 16-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_1min_SQobold_all.hdf       # 3 channels: Bin_L, Bin_R, GRAS_146AE
│   │   ├── S0001_1min_SQobold_bin.wav       # 2 channels, 24-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_1min_SQobold_spl.wav       # 1 channel, 24-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_1min_Zylia_02ch.wav        # 2 channels, 24-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_1min_Zylia_19ch.wav        # 19 channels, 24-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_full_Insta360.mp4          # 4096 x 2048 resolution, min. 40 mbps @ 29.97 fps, audio min. 128 kbps, 2 channels, 16-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_full_SQobold_all.hdf       # 3 channels: Bin_L, Bin_R, GRAS_146AE
│   │   ├── S0001_full_SQobold_bin.wav       # 2 channels, 24-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_full_SQobold_spl.wav       # 1 channel, 24-bit representation, 48000 Hz sampling frequency
│   │   ├── S0001_full_Zylia_02ch.wav        # 2 channels, 24-bit representation, 48000 Hz sampling frequency
│   │   └── S0001_full_Zylia_19ch_part01.wav # 19 channels, 24-bit representation, 48000 Hz sampling frequency. Other soundscapes may have more parts
│   │
│   ├── S####
│   │   └── ...
│   │
│   └── S0062
│       ├── S0062_1min_Insta360.mp4
│       ├── S0062_1min_SQobold_all.hdf
│       ├── S0062_1min_SQobold_bin.wav
│       ├── S0062_1min_SQobold_spl.wav
│       ├── S0062_1min_Zylia_02ch.wav
│       ├── S0062_1min_Zylia_19ch.wav
│       ├── S0062_full_SQobold_all.hdf
│       ├── S0062_full_SQobold_bin.wav
│       ├── S0062_full_SQobold_spl.wav
│       └── S0062_full_Zylia_02ch.wav
│
├── figures                       # Folder containing all reference figures used in the Jupyter Notebook (29 files; all png)
│   ├── all_soundscapes.png
│   ├── ...     
│   └── uniform_circle_points.png 
│
├── results                       # Folder to store reference results & results generated by the Jupyter Notebook
│   └── seed-12-epoch-79-ref.csv
│
├── CITATION.cff
├── README.md
├── lcs.yml
├── lcs-mac.yml
├── metadata.csv
└── predictions.csv

Metadata files

metadata.csv

• lcs_id : unique strings
  • The ID of the soundscape in the LCS dataset. Each ID corresponds to a unique soundscape.
• lcs_description : strings
  • A brief description of the actual location where the soundscape was recorded for the LCS dataset, which we refer to as the "recording location".
  • This may differ from the exact location identified by the Singapore Soundscape Site Selection Survey (S5 study), which we refer to as the "S5 location". This is due to the fact that the S5 location may be physically or publicly inaccessible (e.g., on water).
• s5_type : strings in {"F&E", "C&R", "C&T", "B&L"}
  • The perceptual quadrant in which the soundscape at the S5 location was classified using the modified $k$-means clustering algorithm in the S5 study.
  • Keys:
    • F&E : Full of life and exciting
    • C&R: Chaotic and restless
    • C&T : Calm and tranquil
    • B&L : Boring and lifeless
• s5_id : positive integers in {0, 1, 2, ..., 17, 18}
  • The index of the soundscape at the S5 location for that particular perceptual quadrant in the S5 study.
• s5_latitude : floating point numbers
  • The latitude of the S5 location (i.e., the location identified in the S5 study).
• s5_longitude : floating point numbers
  • The longitude of the S5 location (i.e., the location identified in the S5 study).
• lcs_latitude : floating point numbers
  • The latitude of the recording location (i.e., the actual location where the soundscape was recorded for the LCS dataset).
• lcs_longitude : floating point numbers
  • The longitude of the recording location (i.e., the actual location where the soundscape was recorded for the LCS dataset).
• lcs_date : strings
  • The date on which the soundscape was recorded at the recording location for the LCS dataset, in YYYYMMDD format.
  • All recordings were done from 20220916 to 20230111 (inclusive).
• lcs_time : strings
  • The time corresponding to the start of the full-length recording of the soundscape at the recording location, in 24-hour format.
  • All recordings were started from 0909 to 1855 (inclusive).
• usable_length_bin_minutes : integers
  • The minutes place of the length of the full-length recording made at the recording location.
  • For example, if the full-length recording was 26 minutes and 10.9 seconds long, the value here would be 26.
• usable_length_bin_seconds : integers in {0, 1, 2, ..., 58, 59}
  • The seconds place of the length of the full-length recording made at the recording location.
  • For example, if the full-length recording was 26 minutes and 10.9 seconds long, the value here would be 10.
• usable_length_bin_frames_at_30fps : integers in {0, 1, 2, ..., 28, 29}
  • The frames place of the length of the full-length recording made at the recording location (assuming a frame rate of 30 frames per second).
  • For example, if the full-length recording was 26 minutes and 10.9 seconds long, the value here would be 27.
• usable_length_all_minutes : integers
  • The minutes place of the time after the start of the full-length recording when the ZYLIA recording first desynchronises from the other recording devices.
  • We found out when processing the audio-visual recordings that due to technical issues with the ZYLIA recorder, samples were intermittently dropped (for unknown durations and at random times) in the ZYLIA recordings for some locations. Hence, data beyond this timestamp should not be used if the ZYLIA recording needs to be synchronised with a recording from any other recording device (even though we have made all data available for archival purposes).
  • For example, if the desynchronisation happened at 23 minutes and 1.4 seconds, the value here would be 25.
• usable_length_all_seconds : integers in {0, 1, 2, ..., 58, 59}
  • The seconds place of the time after the start of the full-length recording when the ZYLIA recording first desynchronises from the other recording devices.
  • For example, if the full-length recording was 23 minutes and 1.4 seconds long, the value here would be 1.
• usable_length_all_frames_at_30fps : integers in {0, 1, 2, ..., 28, 29}
  • The frames place of the time after the start of the full-length recording when the ZYLIA recording first desynchronises from the other recording devices (assuming a frame rate of 30 frames per second).
  • For example, if the length of the recording was 23 minutes and 1.4 seconds, the value here would be 12.
• onemin_offset_bin_minutes : integers
  • The minutes place of the time after the start of the full-length recording when the 1-minute excerpt was cropped.
  • For example, if the 1-minute excerpt was cropped 14 minutes and 12.0 seconds after the start of the full-length recording, the value here would be 14.
• onemin_offset_bin_seconds : integers in {0, 1, 2, ..., 58, 59}
  • The seconds place of the time after the start of the full-length recording when the 1-minute excerpt was cropped.
  • For example, if the 1-minute excerpt was cropped 14 minutes and 12.0 seconds after the start of the full-length recording, the value here would be 12.
• onemin_offset_bin_frames_at_30fps : integers in {0, 1, 2, ..., 28, 29}
  • The frames place of the time after the start of the full-length recording when the 1-minute excerpt was cropped (assuming a frame rate of 30 frames per second).
  • For example, if the 1-minute excerpt was cropped 14 minutes and 12.0 seconds after the start of the full-length recording, the value here would be 0.
• insitu_spl_full_L : floating point numbers
  • Exponentially averaged A-weighted sound pressure level (in decibels) over time at the left channel of the B&K Type 4101-B binaural microphone, computed with fast averaging (i.e. with time constant of 125 milliseconds), and then averaged across the entire full-length recording.
  • Filter is designed according to ISO1996-1. This is as recommended in Table D.1 of ISO 12913-3:2019 for soundscape studies.
• insitu_spl_full_R : floating point numbers
  • Exponentially averaged A-weighted sound pressure level (in decibels) over time at the right channel of the B&K Type 4101-B binaural microphone, computed with fast averaging (i.e. with time constant of 125 milliseconds), and then averaged across the entire full-length recording.
  • Filter is designed according to ISO1996-1. This is as recommended in Table D.1 of ISO 12913-3:2019 for soundscape studies.
• insitu_spl_full_LR_eavg : floating point numbers
  • The energetic average of the values in insitu_spl_full_L and insitu_spl_full_R.
  • The energetic average of $x$ and $y$ is defined as $10\lg\left(10^{\frac{x}{10}} + 10^{\frac{y}{10}}\right) - 10\lg(2)$.
• insitu_spl_full_GRAS146AE : floating point numbers
  • Exponentially averaged A-weighted sound pressure level (in decibels) over time at the GRAS 146AE Free-field Microphone, computed with fast averaging (i.e. with time constant of 125 milliseconds), and then averaged across the entire full-length recording.
  • Filter is designed according to ISO1996-1. This is as recommended in Table D.1 of ISO 12913-3:2019 for soundscape studies.
• insitu_spl_1min_L : floating point numbers
  • Exponentially averaged A-weighted sound pressure level (in decibels) over time at the left channel of the B&K Type 4101-B binaural microphone, computed with fast averaging (i.e. with time constant of 125 milliseconds), and then averaged across the 1-minute excerpt of the recording.
  • Filter is designed according to ISO1996-1. This is as recommended in Table D.1 of ISO 12913-3:2019 for soundscape studies.
• insitu_spl_1min_R : floating point numbers
  • Exponentially averaged A-weighted sound pressure level (in decibels) over time at the right channel of the B&K Type 4101-B binaural microphone, computed with fast averaging (i.e. with time constant of 125 milliseconds), and then averaged across the 1-minute excerpt of the recording.
  • Filter is designed according to ISO1996-1. This is as recommended in Table D.1 of ISO 12913-3:2019 for soundscape studies.
• insitu_spl_1min_LR_eavg : floating point numbers
  • The energetic average of the values in insitu_spl_1min_L and insitu_spl_1min_R.
  • The energetic average of $x$ and $y$ is defined as $10\lg\left(10^{\frac{x}{10}} + 10^{\frac{y}{10}}\right) - 10\lg(2)$.
• insitu_spl_1min_GRAS146AE : floating point numbers
  • Exponentially averaged A-weighted sound pressure level (in decibels) over time at the GRAS 146AE Free-field Microphone, computed with fast averaging (i.e. with time constant of 125 milliseconds), and then averaged across the 1-minute excerpt of the recording.
  • Filter is designed according to ISO1996-1. This is as recommended in Table D.1 of ISO 12913-3:2019 for soundscape studies.
• gain_at_insitu_spl_LR : floating point numbers
  • Gain to apply to the 2-channel wave file recorded using the binaural microphone, in order to achieve an L_A,eq of insitu_spl_full_LR_eavg decibels when played back over a pair of Beyerdynamic Custom One Pro headphones, powered by a Creative SoundBlaster E5 soundcard (set at volume 40).
• gain_at_insitu_spl_GRAS146E : floating point numbers
  • Gain to apply to the 2-channel wave file recorded using the binaural microphone, in order to achieve an L_A,eq of insitu_spl_full_GRAS146AE decibels when played back over a pair of Beyerdynamic Custom One Pro headphones, powered by a Creative SoundBlaster E5 soundcard (set at volume 40).
• link_full_youtube : strings
  • The YouTube link to the full-length recording from the Insta360 One R camera.
  • These are 360-degree videos in spherical format corresponding to the files S####_full_Insta360.mp4.
• link_full_insta360 : strings
  • The link to download the full-length recording from the Insta360 One R camera, as stored on our institutional repository DR-NTU (Data).
  • These are 360-degree videos in spherical format corresponding to the files S####_full_Insta360.mp4.
  • May be NIL if the full-length recording is unavailable on DR-NTU (Data).
• link_full_sqobold_all : strings
  • The link to download the 3-channel HDF file containing the full-length recordings from both the B&K Type 4101-B binaural microphone and GRAS 146AE Free-field Microphone, as stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_full_SQobold_all.hdf.
  • Note that "HDF" here refers to "HEAD Data Format", and not the more common "Hierarchical Data Format". Software such as HEAD Companion and ArtemiS may be used to open this file.
• link_full_sqobold_bin : strings
  • The link to download the 2-channel WAV file recorded using the B&K Type 4101-B binaural microphone, as generated from the first two channels of the HDF file at link_full_sqobold_all and stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_full_SQobold_bin.wav.
• link_full_sqobold_spl : strings
  • The link to download the 1-channel WAV file recorded using the GRAS 146AE Free-field Microphone, as generated from the last channel of the HDF file at link_full_sqobold_all and stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_full_SQobold_spl.wav.
• link_full_zylia_19ch_part01 : strings
  • The link to download the first part of the 19-channel WAV file containing the full-length recording from the ZYLIA ZM-1 3rd-order Ambisonic microphone, as stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_full_Zylia_19ch_part01.wav.
  • Some ZYLIA recordings were split into two parts due to file size limitations with the WAV format.
  • May be NIL if the full-length recording is unavailable on DR-NTU (Data).
• link_full_zylia_19ch_part02 : strings
  • The link to download the second part of the 19-channel WAV file containing the full-length recording from the ZYLIA ZM-1 3rd-order Ambisonic microphone, as stored on our institutional repository DR-NTU (Data).
  • These corespond to the files S####_full_Zylia_19ch_part02.wav.
  • Some ZYLIA recordings were split into two parts due to file size limitations with the WAV format.
  • May be NIL if the full-length recording is unavailable on DR-NTU (Data), or if there is no second part.
• link_full_zylia_02ch : strings
  • The link to download the 2-channel stereo WAV file generated from the 19-channel WAV file at link_full_zylia_19ch_part01 (and link_full_zylia_19ch_part02 if a second part exists) by the ZYLIA ZR-1 Portable Recorder.
  • These correspond to the files S####_full_Zylia_02ch.wav.
• link_1min_youtube : strings
  • The YouTube link to the 1-minute excerpt of the full-length recording from the Insta360 One R camera.
  • These are 360-degree videos in spherical format corresponding to the files S####_1min_Insta360.mp4.
• link_1min_insta360 : strings
  • The link to download the 1-minute excerpt of the full-length recording from the Insta360 One R camera, as stored on our institutional repository DR-NTU (Data).
  • These are 360-degree videos in spherical format corresponding to the files S####_1min_Insta360.mp4.
• link_1min_sqobold_all : strings
  • The link to download the 3-channel HDF file containing the 1-minute excerpt of the full-length recordings from both the B&K Type 4101-B binaural microphone and GRAS 146AE Free-field Microphone, as stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_1min_SQobold_all.hdf.
  • Note that "HDF" here refers to "HEAD Data Format", and not the more common "Hierarchical Data Format". Software such as HEAD Companion and ArtemiS may be used to open this file.
• link_1min_sqobold_bin : strings
  • The link to download the 2-channel WAV file recorded using the B&K Type 4101-B binaural microphone, as generated from the first two channels of the HDF file at link_1min_sqobold_all and stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_1min_SQobold_bin.wav.
• link_1min_sqobold_spl : strings
  • The link to download the 1-channel WAV file recorded using the GRAS 146AE Free-field Microphone, as generated from the last channel of the HDF file at link_1min_sqobold_all and stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_1min_SQobold_spl.wav.
• link_1min_zylia_19ch : strings
  • The link to download the 19-channel WAV file containing the 1-minute excerpt of the full-length recording from the ZYLIA ZM-1 3rd-order Ambisonic microphone, as stored on our institutional repository DR-NTU (Data).
  • These correspond to the files S####_1min_Zylia_19ch.wav.
  • Note that for Bishan Harmony Park (S0054), the ZYLIA recordings cut off earlier than the SQobold recordings so there is no corresponding ZYLIA track for the 1-min excerpt of S0054. For now, we use as a placeholder a 19-channel repeat of the SQobold SPL track for the Zylia ambisonic track (i.e., S0054_1min_Zylia_19ch.wav contains the data in S0054_1min_SQobold_spl.wav replicated over 19 channels).
• link_1min_zylia_02ch : strings
  • The link to download the 2-channel stereo WAV file generated from the 19-channel WAV file at link_1min_zylia_19ch by the ZYLIA ZR-1 Portable Recorder.
  • These correspond to the files S####_1min_Zylia_02ch.wav.
  • Note that for Bishan Harmony Park (S0054), the ZYLIA recordings cut off earlier than the SQobold recordings so there is no corresponding ZYLIA track for the 1-min excerpt of S0054. For now, we use as a placeholder the SQobold binaural track for the Zylia stereo track (i.e., S0054_1min_Zylia_02ch.wav and S0054_1min_SQobold_bin.wav contain the same data).

predictions.csv

• lcs_id : unique strings
  • The ID of the soundscape in the LCS dataset. Each ID corresponds to a unique soundscape.
• start_idxs : integers
  • The time (in seconds) of the start of the 30-second window of the binaural recording of the soundscape (made at the recording location) used as input to the ensemble model pre-trained on the ARAUS dataset (preprint available on arXiv).
• isopl_mean :
  • The mean of the predictions of the ensembled models for ISO Pleasantness (as defined in ISO 12913) for the 30-second window starting at start_idxs seconds.
• isoev_mean :
  • The mean of the predictions of the ensembled models for ISO Eventfulness (as defined in ISO 12913) for the 30-second window starting at start_idxs seconds.

Recording procedure

For every location where the soundscape was recorded, we followed a similar procedure:

1. Press the buttons to start recording for the Insta360 camera, ZYLIA ambisonic microphone, and SQobold data acquisition system (order may be different for each recording).
2. Make a synchronisation clap that it is visible from the Insta360 camera and audible from all microphones (this is used to synchronise all recordings in the post-processing phase).
3. Walk away from the recording setup, such that we are not visible in the Insta360 video recording, for the duration of the full-length recording.
4. Return to the recording setup and make a second synchronisation clap that is visible from the Insta360 camera and audible from all microphones (this marks the end of the recording).
5. Press the buttons to stop recording for the Insta360 camera, ZYLIA ambisonic microphone, and SQobold data acquisition system (order may be different for each recording).

In the form of a timeline, let us define the following values, in units of minutes:

• $\alpha$ : Time when Insta360 recording was started
• $\beta$ : Time when SQobold recording (for both B&K Type 4101-B and GRAS 146AE microphones) was started
• $\gamma$ : Time when Zylia recording was started
• $\delta$ : Time when synchronisation clap was performed
• $\delta^+$ : Time when we (researchers) are fully out of frame in Insta360 recording
• $\varepsilon$ : Time when the 1-minute excerpt of the soundscape starts
• $\zeta$ : Time when the 1-minute excerpt of the soundscape ends
• $\eta$ : Time when the Zylia recording first desynchronises from the SQobold and Insta360 recordings
• $\theta$ : Time when we (researchers) become visible in Insta360 recording
• $\kappa$ : Time when Zylia recording was stopped
• $\lambda$ : Time when Insta360 recording was stopped
• $\mu$ : Time when SQobold recording was stopped

Then, we have something like this:

         │<------ insta360_offset ------->│<------------ usable_length_bin ------------->│
         │                                │                                              │
Insta360 ├█████████████████████████████████████████████████████████████████████████████████████████████┤
         │            │<-- zylia_offset-->│<---------- usable_length_all ---------->│    │             │
         │            │                   │                                         │    │             │
ZYLIA    │            ├█████████████████████████████████████████████████████████████████████████┤      │
         │      │<---- sqobold_offset --->│<-- onemin_offset_bin -->│<-- 1 -->│     │    │      │      │
         │      │     │                   │                         │         │     │    │      │      │
SQobold  │      ├██████████████████████████████████████████████████████████████████████████████████████████████┤
         │      │     │         │         │                         │         │     │    │      │      │       │
         alpha  beta  gamma     delta     delta+                    epsilon   zeta  eta  theta  kappa  lambda  mu

│─┌┐└┘┬┴┤├┼

Acknowledgements

We would like to express our heartfelt gratitude to the management teams at the following locations for their assistance with this project:

• Changi Airport (S0011, S0012)
• National Parks Board, Bukit Timah Nature Reserve (S0055)
• National Parks Board, Jurong Lake Gardens (S0056)
• National Parks Board, Singapore Botanic Gardens (S0029)
• National Parks Board, Sungei Buloh Wetland Reserve (S0018, S0019)
• Raffles Marina (S0032, S0033)
• Singapore Zoo (S0041)

Terms of use

All data in this dataset, including code, metadata, and audio-visual recordings, are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike license. You may view the license terms at https://creativecommons.org/licenses/by-nc-sa/4.0/. By accessing, viewing, or downloading any data in this dataset, in full or in part, it is assumed that you have read and agreed to the license terms.

THIS DATASET IS FOR NONCOMMERCIAL USE BY ACADEMIC OR NON-PROFIT ORGANIZATIONS ONLY. COMMERCIAL USAGE OF THIS DATASET, IN FULL OR IN PART, IS STRICTLY PROHIBITED.

Version history

• 0.0.0 : Initial release