Room Impulse Response estimation in Noisy Environments using Music as an Excitation Signal
(2026) In Master's Theses in Mathematical Sciences MASM02 20261Mathematical Statistics
- Abstract
- This thesis investigates non-intrusive Room Impulse Response (RIR) estimation using music as an excitation signal in an acoustic environment with background noise present. The work studies the structure of room impulse responses, sparse background noise, babble noise, and music excitation signals in the context of RIR estimation. Building on the recently proposed AnyRIR estimator, which explored similar structures, weighted estimation methods incorporating background noise covariance are investigated under different background noise conditions.
Robustness to sparse background disturbances is achieved through the Huber objective function formulation, reducing sensitivity to sporadic interference and resulting in high restored speech... (More) - This thesis investigates non-intrusive Room Impulse Response (RIR) estimation using music as an excitation signal in an acoustic environment with background noise present. The work studies the structure of room impulse responses, sparse background noise, babble noise, and music excitation signals in the context of RIR estimation. Building on the recently proposed AnyRIR estimator, which explored similar structures, weighted estimation methods incorporating background noise covariance are investigated under different background noise conditions.
Robustness to sparse background disturbances is achieved through the Huber objective function formulation, reducing sensitivity to sporadic interference and resulting in high restored speech intelligibility. In babble noise conditions, incorporating background noise covariance matrix as weight within the Huber loss function increases accuracy of the estimated RIR. However, this did not lead to substantial improvements in restored speech intelligibility, suggesting that intelligibility is limited more by remaining structured babble interference than by RIR estimation accuracy itself. Results show that covariance weighting alone with weighted least squares is insufficient for RIR estimation in the presence of babble noise, producing higher normalized mean squared error (NMSE) of the estimated RIR compared to covariance weighted Huber estimation. (Less) - Popular Abstract
- When criminals want to discuss sensitive information about some high-risk activities, they are likely to do it in person, rather than over the phone, eliminating digital trails, making it significantly harder for law enforcement to intercept or reconstruct the details of a crime. However, a microphone recording of a room they are talking in can be obtained and used as incriminating evidence. Therefore, the criminals are likely to have their conversation in a loud place like a café, a bar or some live event where loud music and background noises mask their conversation. Due to music recognition software like Shazam, the song playing in the room can be identified and subtracted from the recorded signal. But if only the music track is... (More)
- When criminals want to discuss sensitive information about some high-risk activities, they are likely to do it in person, rather than over the phone, eliminating digital trails, making it significantly harder for law enforcement to intercept or reconstruct the details of a crime. However, a microphone recording of a room they are talking in can be obtained and used as incriminating evidence. Therefore, the criminals are likely to have their conversation in a loud place like a café, a bar or some live event where loud music and background noises mask their conversation. Due to music recognition software like Shazam, the song playing in the room can be identified and subtracted from the recorded signal. But if only the music track is removed, the recorded signal will still have audible music playing, because the recording still contains the reverberant components of the song due to reflections of the music signal within the room. To account for the reflections, Room Impulse Response (RIR) has to be estimated and also be removed from the recorded signal alongside the music track.
The estimation procedure involves finding a RIR such that the difference between the recorded signal and the music signal filtered through the RIR is as small as possible. However, the recorded signal contains more than the music signal and its reflections within the room, it also includes the conversation of interest, other people talking and other background noises like utensils clanking, chairs scraping against the floor, etc. These background sounds lead to large differences, errors between recorded signal and RIR with music signal at the time of their occurrence, consequently, bias is introduced in the estimated room impulse response. Hence, robustness to transient bursts of sound and continuous background noise of people constantly speaking throughout the signal is applied. To make the estimation less sensitive to sparse noises and the speech we want to recover, the Huber loss method is used, which reduces the influence of large errors above a chosen threshold. The effect of continuous noise of other people talking – babble noise is suppressed by estimating the temporal dependencies between the nearby points of the babble noise and using it as weights in the minimization problem – covariance weighting method. Hence, three methods are tested: the Huber loss method, covariance weighting method and a hybrid of the two.
The results show that in sparse background noise setting, Huber loss method provides best RIR accuracy and recovered speech intelligibility results. While, when there is continuous background noise obstructing speech of interest, the covariance weighted Huber loss method performs best in estimation accuracy. However, increased accuracy in RIR estimate, did not lead to significant improvements in recovered speech intelligibility, indicating that the speech intelligibility is limited more by the remaining obstructing babble noise. The method with only covariance weighting yielded the poorest estimate accuracy and speech intelligibility in both sparse and babble noise conditions. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9231367
- author
- Serpatauskaite, Monika LU
- supervisor
- organization
- course
- MASM02 20261
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Room Impulse Response, Huber Loss, Covariance Weighting, Weighted Least Squares, Sparse Signals, Image Source Method, Ray Tracing
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUNFMS-1404-2026
- ISSN
- 1404-6342
- other publication id
- 2026:E38
- language
- English
- id
- 9231367
- date added to LUP
- 2026-06-05 09:38:31
- date last changed
- 2026-06-05 09:38:31
@misc{9231367,
abstract = {{This thesis investigates non-intrusive Room Impulse Response (RIR) estimation using music as an excitation signal in an acoustic environment with background noise present. The work studies the structure of room impulse responses, sparse background noise, babble noise, and music excitation signals in the context of RIR estimation. Building on the recently proposed AnyRIR estimator, which explored similar structures, weighted estimation methods incorporating background noise covariance are investigated under different background noise conditions.
Robustness to sparse background disturbances is achieved through the Huber objective function formulation, reducing sensitivity to sporadic interference and resulting in high restored speech intelligibility. In babble noise conditions, incorporating background noise covariance matrix as weight within the Huber loss function increases accuracy of the estimated RIR. However, this did not lead to substantial improvements in restored speech intelligibility, suggesting that intelligibility is limited more by remaining structured babble interference than by RIR estimation accuracy itself. Results show that covariance weighting alone with weighted least squares is insufficient for RIR estimation in the presence of babble noise, producing higher normalized mean squared error (NMSE) of the estimated RIR compared to covariance weighted Huber estimation.}},
author = {{Serpatauskaite, Monika}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Room Impulse Response estimation in Noisy Environments using Music as an Excitation Signal}},
year = {{2026}},
}