Acoustic Environment Classification

Method-
Hidden Markov Model (HMM) classifier.
-to test the feasibility of acoustic environment recognition in mobile devices
-Acoustic environments can be accurately classified using sound samples
acquired from widely available consumer devices.

Section 2-
described related work in environmental noise recognition, video
indexing, and contextaware computing.
Foote [1999] and Wold et al. [1996] reviewed audio classification systems
using multiple features and a variety of classification techniques including
static modeling and dynamic modeling.
Content-based audio classification and retrieval research has been
mainly based on speech and music, focusing on searching and indexing.
Much work has been done on speech recognition, word spotting, speaker
identification, speech interface, and music information retrieval.
MFCCs are the most common feature representation for nonspeech
audio recognition. Peltonen et al. [2002] implemented a system for
recognizing 17- sound events using 11 features individually and obtained
best results with the MFCC features.
A variety of approaches have been used to distinguish speech from
music[Vendrig et al. 2003; Wu et al. 2003; Scheirer and Slaney 1997].
Cai et al. [2003] propose an audio classification based on a combination
of statistical features and labels describing the energy envelope,
harmonicity, and pitch contour for each sample in a database containing
approximately 600 high quality samples ranging from 1 second to 30
seconds in duration.



Sections 3&4-
- classification experiments are described
describe a HMM framework for classifying a range of different acoustic
environments. Classification is based on combining digital signal
processing (DSP) technology with pattern recognition methods that have
been central to progress in automatic speech recognition [Huang et al.
2001] over the last 20 years.
important differences between recognizing speech and identifying
acoustic environments: notably, speech is produced from a single source,
constrained to a single location, which has limitations on the character of
sound it can produce. An acoustic environment, however, has none of
these constraints and is a complex sound made up of a mixture of
different events.
four phases in the construction of a set of acoustic environment models
suitable for classification.

1. acoustic environment database capture (via portable recording
devices)
2. preprocessing (digitization,segmentation, and labeling),
3. feature extraction,
4. training and testing a set of HMMs.

LOW BANDWIDTH EXPERIMENTS
To deploy context tracking applications on mobile devices, three issues must be
addressed. First, since users are limited to low bandwidth connections because
of communications infrastructure and small footprint devices (mobile phones
and PDAs), we need to implement and test a HMM-based acoustic
environment classifier that is appropriate for low-bandwidth connections.
Second, the classifier is designed for real-time processing devices which typically
have only low-quality recording facilities. Third, the environments to be classified
are designed to cover a range of everyday situations where users would be
accessing information services from mobile devices.

Low-Bandwidth Data Collection
Data collection was made using an MP3 player/recorder (with the resulting
audio sampled at 8kHz.) attached to the strap of a shoulder bag as the
recording device to capture the acoustic environments from typical daily
activities
Low Bandwidth Classification Experiments
-examines the accuracy of acoustic environment classification using the low
bandwidth data. From each of the 3-second duration audio files, MFCC vectors
(augmented with velocity and acceleration derivatives) were extracted as
described in Section 3. A set of 12 HMMs, corresponding to the 12 environments,
was trained from the 1,200 sound examples in Dataset 1. The HMMs used the 9
stage, left-to-right topology that gave the best results in the experiments
discussed in Section 3. Dataset 2 was then used for testing which gave an
overall classification accuracy of 96% on the 1,200 samples.
- The recognition accuracy of individual environments ranged from 81% to 100%,
with the building site, car (city), office, presentation, pedestrian street, and
supermarket, giving 100% classification accuracy for the 100 examples of each
tested. The worst performance was obtained for identifying bus noisewhich
achieved only 81% accuracy. However, the confusions of bus noise were
reasonable because the incorrectly classified samples were all recognized as
other vehicles, 13% as car (city), 3% as car (highway) and 3% as train. Other
reasonable confusions can also be seen; for example, 11% of shopping mall
samples were incorrectly classified as supermarket. This is due to the acoustic
similarity of the two environments, that is, sounds from trolleys, people, tills, etc


Section 5-
Described about the adaptive learning approach
To be useful in realistic situations, the acoustic environment classification must be
adaptable to changes in the character of environments and to the introduction
of new environments. It is also desirable to know the confidence of the
environment classification. Here, we describe a confidence measure for our
classifier and its use in a set of continuous learning experiments. The common
notion of continuous learning has been described by Vega and Bressan [2003]
who have shown that an algorithm which gives a good initial performance will
improve its performance if it is given its own classified data as further training
examples.



The experimental results show that wrongly identified samples are most often
from similar environments to their correct classification; this is particularly true of
the fine-grained classification used in parts of the RWCP single-sound database.
We have developed a confidence measure which allows us to classify sounds in
a hierarchical model at the most specific level consistent with high confidence.
The construction of semantically-useful cohesive acoustic environment and
sound event hierarchies has several open issues which are beyond the scope of
this work. These results have clearly shown the potential of using the acoustic
environment as a context indicator. The classifier gives good results over a wide
range of sample qualities and environments and an adaptive learning strategy
allows the model to learn new environments and recognize the specific
environments of an individual.We have described the models use in a client-
server context-aware prototype tracker application and demonstrated the
feasibility of capturing environmental noise using current PDAs, although there
are several implementation issues that need to be addressed before integrated
systems can be deployed. This work shows that the acoustic environment is a
rich source of context information which can be recognized with a high degree
of accuracy and can be used as a good indicator of current activity.






Akustik di dalam kelas
Masalah, Kesan dan Cara Mengatasi
Problems
-reflected sound can enhance direct sound, making it easier to hear speech. In
order for reflected sound energy to be maximally beneficial in a classroom, the
room should be designed to have an appropriate combination of reflective,
absorptive, and diffusive surfaces. To primarily reflect sound, a surface must be
flat curved, as well as stiff and hard. Additionally, it must be several times greater
in length or width than the wavelengths of the incident sounds. To primarily
absorb sound, a surface can be of any shape, but it must be porous,
diaphragmatic, or resonant. Surfaces that are irregular so that sound scatters in
many directions will primarily diffuse sound (D'Antonio, 1989).
-Rooms designed primarily for listening to speech do not need a great deal of
sound diffusion (Brook, 1991), but the reflective and absorptive characteristics of
the room are critically important.
-When not controlled properly, the reflection and absorption of sounds can
interfere with classroom listening. For example, some of the most frequently used
absorbent materials in school classrooms are acoustic ceiling tile and carpet on
foam rubber padding. These materials absorb high-frequency sounds much
better than low-frequency sounds. When the ceiling and floor are treated with
these materials, consonant sounds, which are primarily high frequency, are not
reflected well, and speech intelligibility is degraded (Everest, 1989; Harris, 1955).
-There are several types of reflected sounds that occur within a room, including
1. room modes
-Room modes are room resonances that cause
standing waves with "dead" and "live" spots and degrade
speech within the frequency region of 80-300 Hz (Everest,
1991b).
2. Echoes
-Echoes are reflected sound waves that are delayed and sufficiently
intense to be distinct from the original sound source. They are likely to
occur in classrooms where the back wall has a hard surface. Echoes are
not usually heard as discrete phenomena, but they influence the level,
quality, and intelligibility of the sound (Everest, 1989,1991b).
3. reverberation.
-is aprolongation or repeated reflection of sound.

-Although room modes, echoes, and reverberation can affect speech
intelligibility in a classroom, the most serious acoustical problem in school
classrooms is excessive noise, which masks the teacher's speech.

Impact
-The effects of noise and reverberation on speech recognitionscores of school
children who have normal hearing and school children who are hard of hearing
were studied by Finitzo-Hieber & Tillman (1978).
-Thus, (1) both noise and reverberation degraded speech recognition, and the
combination of the two was especially detrimental.
-(2)the speech signal in the back of the room was degraded when compared
to that in the front of the room.
- (3)School beginners have problems learning in classrooms because their
speech, language, and listening skills have not matured, making it difficult for
both teachers and young students to communicate with each other (Durkin,
1986; Edwards & Mercer, 1986; Nippold, 1988; Romaine, 1984).
-(4) When a student uses English as a second language, poor classroom
acoustics may further compromise the child's ability to learn (Williams & Capizzi-
Snipper, 1990).
-(5)Children with learning deficits may also be negativelyaffected by poor
classroom acoustics because they frequently have phonologic, semantic,
syntactic, and pragmatic language disorders that limit their communication
competencies (Bos, 1988; Lansky & Chapandy, 1976; Lupert, 1981; Lyon &
Risucci, 1988; Scruggs, 1988; Simon, 1985; Wiig & Semel, 1976).
- When(6) students cannot listen effectively in school, they are more likely to
have difficulty staying on task, and discipline and cooperation are difficult to
maintain (Gallup, 1986). In noisy classrooms, many teachers make vocal
adjustments in order to maintain classroom control and to reach desirable S/N
ratios (Ray, 1990). Some teachers have inherently strong voices and are able to
project their voices for long periods of time without tiring. Other teachers,
however, have relatively weak voices and become stressed when forced to
raise their voice level. (7)When teachers use their voices to compensate for high
noise levels, they are also more tired at the end of the school day (Berg, 1993).
Sometimes, classroom noise levels are too high for any teacher to overcome
with vocal effort, and abuse and damage to a teacher's vocal mechanism may
result (Child & Johnson, 1991).

Solution
In order to improve the listening, learning, and teaching environment in a
classroom, several different approaches have been used.
1. Noise Control
a noise control plan that describes current noise levels, provides general and
specific objectives, determines the person responsible for implementing this
plan, and describes evaluation procedures (Berg, 1987, 1993). When feasible,
architects, engineers, educational audiologists, and other school personnel
should collaborate to provide for reduced noise levels in the school (Berg, 1993).

2. Signal Control Without Amplification
In order to improve the acoustical environment of aclassroom, one or more of
the following suggestion can be implemented:

* Make all surface areas that do not provide useful
reflection absorbent, and, conversely, do not cover any
useful reflective surfaces (Brook & Uzzle, 1987).

* Make the ceiling and side walls reflective surfaces in
order to increase the signal intensity.

a Install carpet on the floor to cover a useless reflecting
surface and reduce noise (however, as suggested
earlier, the combination of ceiling treatments and floor
treatments needs to researched more fully).

* Place absorbent panels on the back wall in order to
prevent echoes and reduce reverberation.

3. Use of Individual Amplification Systems
These systems include personal hearing aids, nonprescriptive amplifiers, personal
FM systems, loop FM systems, and infrared systems. Each type of system has
distinctive components, features, roles, and comparative advantages and
limitations (Finitzo, 1988; Radcliffe, 1991; Sandlin, 1991). To ensure effectiveness
for each student, a particular system must be carefully selected,evaluated, and
maintained (Garstecki, Wilber, Stein, & Pasa, 1991; Mueller, Hawkins & Northern,
1992; Ross, 1992). If these steps are taken, hearing aids will enable students with
hearing losses to hear teachers at close distances under quiet conditions, and
personal FM systems, loop systems, and infrared systems will assist students with
hearing losses and learning difficulties in hearing teachers
across classrooms under poor acoustical conditions.

4. Use of Sound Field Amplification Systems
Perhaps the most cost-effective and acceptable technologyfor facilitating
learning in a typical school classroom is the use of a sound field FM system (Ray,
1990; Sarff, Ray,& Bagwell, 1981). This type of FM system has loudspeakers that
allow groups or an entire class of students to listen to the teacher's voice at an
improved S/N ratio, no matter where the teacher is in the classroom (Berg, 1987;
Flexer,1992). Compared to other amplification systems used in classrooms, sound
field FM systems reach more students and are subject to the least amount of
downtime as a result of malfunction (Anderson, 1989).






















Accommodating Students with Hearing Loss in a Teacher of the Deaf/Hard of
Hearing Education Program

This paper describes the challenges faced at the postsecondary level for a
teacher education program in which every year, a significant number of
students are deaf or hard of hearing.
-Consideration needed to be given to modifying all aspects of the learning
environment, including improvement of classroom acoustics and lighting,
addition of assistive listening devices, provision of captioning and/or sign
language interpreting, and adapting new technologies for online learning.
-Studies of postsecondary classroom acoustics have shown similar results to
those of elementary classrooms, indicating reverberation times and noise levels
consistently exceeding recommended values (Hodgson, 1999; Kelly & Brown,
2002; Woodford, Pritchard, & Jones, 1998). Improving classroom acoustics
through structural modifications and the addition of sound absorptive materials
would be the preferred solution.
-Although most sound field studies have focused on elementary age children,
other studies have indicated sound field amplification to be beneficial in
postsecondary level classrooms, with an improvement in speech recognition
scores of up to 37% in classrooms with poor listening conditions (Larsen,Vega, &
Ribera, 2008). Crandell, Charlton, Kinder and Kreisman (2001) found adults
demonstrate better ability to understand sentence material in background noise
with sound field amplification than without.
-While notetakers are helpful, providing a listening environment conducive to
listening would intuitively seem to be a more effective strategy, and the
opportunity to do this arose at York University first in 1989, and again in 2006.
-The faculty at York University included three teachers of the deaf and hard of
hearing with many years experience teaching in schools for the deaf, self
contained classes, and itinerant roles (one also has sign language interpreter
qualifications) and an educational audiologist. These faculty members were
able to discuss the issues of accommodation with the architect and physical
plant staff.
-The classroom selected for the renovations consisted of a standard classroom
with high, pyramidal-shaped ceilings made of concrete leading to four skylights.
All classes for the program are now conducted in this room; a benefit for
students with hearing loss who do not have the added stress of travelling
between classes and coping with different acoustical environments.
-Preferential seating is always a key component to any list of strategies provided
to teachers to facilitate understanding in the classroom for students with hearing
loss. Leavitt & Flexers (1991) research clearly indicated the effects of distance
on speech perception in classrooms and underscored the need to consider
where students with hearing loss are seated in a classroom relative to the
instructor. Consideration was given to seating arrangements so that students
with hearing loss would always be able to hear clearly and have a clear view of
the instructors, classmates, and ASL interpreter (if needed).
-Lighting is a key issue for students with hearing loss to ensure effective use of
speechreading, to ensure appropriate access to sign language interpretation,
and to ensure that the captioning displayed on the screen is clear and legible.
-