Tampere University Rotated Circular Array Impulse Response Dataset

A dataset of impulse responses collected from different rooms with varying characteristics in Tampere University , Hervanta campus, Finland.

Why do we need a new IR dataset ?

• Supervised DNNs require lots of labeled training data. The existing datasets do not completely address the data requirements for moving sound source(localization) cases.
• Existing IR datasets are collected with sparse microphone and loudspeaker positions.

Salient features of the TUNI-RCAIR dataset

• The moving source condition is replicated by moving the microphone array, instead of moving the sound source.
• Data is collected by rotating the circular array for every five degrees (fine resolution) and capturing IRs. This setup is decided to address the data requirements.

Comparison with some of the existing datasets.

Dataset	Rooms	Distance Range(m)	Mic/Speaker positions	Angular Resolution	Angle Range
AIR	4	0.5 - 10.2	18	NA	NA
RWCP	5	2	25	20 °	50 ° - 130 °
DCASE	5	1 - 2	504	10 °	-40 ° - 40 °
ACE	7	14	NA	NA	NA
TUNI-RCAIR	7	0.7 - 2.1	777	5 °	-90 ° - 90 °

The above table compares some of the existing IR datasets with the TUNI-RCAIR dataset. A detailed comparison is available in the paper published. TUNI-RCAIR stands out in terms of fine angle resolution and, variation in the microphone/loudspeaker positions.

Data collection setup

• Source - Genelec 810 A loudspeaker.




• Receiver - Microphone array consisting of seven DPA 4060 microphones, is placed on a plastic slab, mounted over a tripod. The plastic slab is attached to Raspberry Pi step motor. The microphone array setup is similar to that of UMA-8, used in smart speakers such as Amazon echo.

Data collection environments

Impulse responses are collected from seven different rooms, varying in physical dimensions, as well as acoustic properties such as reverberation time. The distance between the source and the receiver is varied thrice and the IRs are collected. 777 IRs are collected, 111 IRs from each room. For a particular distance between the source and the receiver, 37 unique IRs are generated.

1. Audio Lab - A studio-like setup with some furniture and lights. It falls under the category of low reverberation but not as low as an anechoic chamber.




2. Office Room-1 - A typical work environment in the university campus. This work space contains tables, sofa, and chairs.




3. Office Room-2 - Another type of office setup, with different dimensions and acoustic characteristics than Office Room-1.




4. Small Meeting Room - One of the smallest places used for data collection. It is highly reverberant in nature. It contains a table and cupboards.




5. Large Meeting Room1 - The Large Meeting Room is bigger than other rooms from which data is collected. Contains a table, a few chairs and sofas.




6. Glass Corridor - IRs are collected from one of the staircases of the university premises. It is highly reverberant, the most among the spaces involved in data collection.




7. Anechoic Chamber - IRs are also collected from an anechoic chamber. It is the least reverberant of all the spaces involved in data collection.

Comparison of spaces used for data collection in terms of physical and acoustic properties.

Room	ID	Dimensions(m)	Volume(m3)	T-60(ms)
Audio Lab	ALB	(2 * 2.3 * 2.6)	11.96	270
Office Room-1	ORA	(3.9 * 4.9 * 3.2)	61.15	486
Office Room-2	ORB	(2.7 * 5.1 * 3.2)	44.06	314
Small Meeting Room	SMR	(2.5 * 3.7 * 2.5)	23.13	416
Large Meeting Room	LMR	(6.6 * 5.2 * 3.2)	109.82	319
Glass Corridor	GCO	(7.3 * 2.7 * 3.4)	67.01	1700
Anechoic Chamber	AEC	(7.8 * 7.8 * 7.8)	474.56	60

TUNI-RCAIR dataset is now available in zenodo. Click on this link to access the dataset

The paper corresponding to the dataset is accepted in the EUSIPCO 2021 will be available after the proceedings of the conference