[ pdf | ps ] In Proc. ICASSP 2005, Vol I, pp. 753--756, March 2005, (Philadelphia, PA) © IEEE 2005.
A cluster-voting scheme is described which takes the output from two speaker diarisation systems and produces a new output which aims to have a lower speaker diarisation error rate (DER) than either input. The scheme works in two stages, firstly producing a set of possible outputs which minimise a distance metric based on the DER and secondly voting between these alternatives to give the final output. Decisions where the inputs agree are always passed to the output and those where the inputs differ are re-evaluated in the final voting stage. Results are presented on the 6-show RT-03 Broadcast News evaluation data, showing the DER can be reduced by 1.64% and 2.56% absolute using this method when combining the best two Cambridge University and the best two MIT Lincoln Laboratory diarisation systems respectively.
Speaker diarisation is the task of automatically segmenting
audio data and providing speaker labels for the resulting
regions of audio.
This has many applications such as enabling speakers to be tracked through
debates, allowing speaker-based indexing of databases, aiding speaker
adaptation in speech recognition and improving readability of transcripts.
The speaker labels produced are 'relative' (such as 'spkr1')
in that they show which segments of audio were spoken by the same speaker,
and do not attempt to give a true identity (such as 'David Koppel')
of the speaker.
The Rich Transcription diarisation evaluations[1, 2] provide
a framework to analyse the performance of such systems on Broadcast News
(BN) data. A Diarisation Error Rate (DER) is defined which considers the
sum of the missed, false alarm and speaker-error
rates after an optimal one-to-one mapping of reference and hypothesis
speakers has been performed. (This mapping is necessary to associate
the 'relative' speaker labels from the hypothesis to the
'true' speaker labels in the reference, and is chosen to maximise
the sum over all reference speakers of the time that is jointly attributed to
both the reference and the corresponding mapped hypothesis
speaker. See  for more details).
In general the error tends to be dominated
by the speaker-error components, and in particular, since the DER is
time-weighted, decisions about the predominant speakers can have a
very large effect. Therefore it is desirable to be able to 'check'
such decisions, perhaps by combining information from different sources.
Several methods of combining aspects of different diarisation systems
tried, for example the 'hybridization' or 'piped' CLIPS/LIA systems
of [3, 4] and the
'Plug and Play' CUED/MIT-LL system of  which both
combine components of different systems together. A more integrated merging
method is described in , whilst 
describes a way of using the 2002 NIST speaker segmentation error metric
to find regions in two inputs which agree and then uses these to
train potentially more accurate speaker models.
These systems are interesting but tend to place some restriction on the
systems being combined which we would like to remove.
This paper describes a cluster-voting scheme which takes the output from any two different diarisation systems as input and tries to produce a new output which is better than either input. This is achieved by reproducing the decisions on which the inputs agree whilst allowing an external judge to decide in the case of conflict. The paper is arranged as follows, section 2 describes the theory behind the scheme, section 3 describes the data used in the experiments, section 4 gives the experimental results and conclusions are offered in section 5.
The cluster voting scheme is illustrated in Figure 1. The process is divided into two steps. Firstly the inputs are compared and the non-conflicting parts are passed straight to the output. A Cluster Voting Metric (CVM) based on the DER is then used on the remaining data to find sets of 'optimal' speaker labellings under the CVM and these alternatives are output in the Cluster Voting Output Set (CVOS). The second stage consists of choosing the final output from the CVOS using an external judge. These stages are discussed further in sections 2.1 and 2.2 respectively.
The stages of the CVOS generation are as follows:
These steps are explained in more detail below and illustrated in Figure 2.
Firstly a common base segmentation is made by dividing up the data as
necessary to ensure there is no speaker change for either input in any base
segment. [Diarisation systems are assumed not to output overlapping speakers.
] This simplifies the calculation of overlap
time between the different clusterings, since each segment can be
represented by its duration and input speaker-ids
alone. [For both experiments reported in this paper the two inputs were generated using different clustering schemes after the
Cambridge or MIT-LL base segmentation respectively, so this stage was
Resegments are then formed
by associating groups of base segments together which have a common speaker
label in input-1 and a common speaker label in input-2. This is
effectively the same as making new speaker labels which are the
concatenation of the two input labels for each segment, and accumulating
the durations for each of these new resegment speaker labels.
No other restriction (such as one of
temporal adjacency) is used to define the resegments, and
no decision is made about the final speaker label
of their base-segments other than to ensure they will have the same
speaker-id in the final output.
This usually significantly reduces
the number of 'segments' to be considered - thus reducing the
computational complexity. For example, this reduces the number of 'segments'
from 869 to 162 in Expt. 1 (section 4.1)
and from 557 to 303 in Expt. 2 (section 4.2).
Any resegments which are non-conflicting (in that
there is no base segment which has the same input speaker-id as
the resegment for either input, but which is not itself in the resegment)
are passed directly to the output with a single unique speaker-id for the
resegment. This reduces the complexity further (removing 71 out of 162
resegments in Expt. 1 and 48 out of 303 in Expt. 2).
The speaker-ids of some groups of resegments may be independent of other
groups. For example, if the clustering has been done gender
dependently, there will never be any overlap between male and female
speaker-ids and thus these two groups can be treated separately.
It is possible to find these independent 'supergroups'
of resegments automatically (see  for details)
and subsequently processing them separately reduces the computational
complexity further. For the example in Figure 2 this reduces
the number of all possible clusterings of resegments from Bell(6)=203 to
Finally, sets of speaker labellings are formed which minimise the
Cluster Voting Metric (CVM) for each supergroup. This metric is the
sum of the DERs from the output to both inputs and reduces to the same
as maximising the sum of the overlap between the output speaker labels and
the speaker labels of both inputs under the optimum one-to-one speaker
mappings performed in the DER calculation, if there is a common input
segmentation. When the difference between the inputs is relatively small,
it is possible to generate all possible speaker labels for the resegments
and score them to find the CVOS exhaustively (see 
but it is also possible to generate the CVOS members directly from
the optimum speaker mapping between the two inputs.
Mapping the inputs leads to a speaker mapping which is either
'good' (the input-1 speaker
is mapped to the input-2 speaker), 'bad' (the input-1 speaker
is mapped to a different input-2 speaker) or 'null' (either
the input-1 or the input-2 speaker is not mapped). A single unique speaker-id
can be assigned to each 'good' mapping. A 'bad' mapping produces two
alternatives, namely those corresponding to the mapped input-1 id or
the input-2 id. The 'null' mappings are more complicated but usually give rise
to two alternatives. The first is from the mapped input-id which is not null,
whilst the second is either a new (unseen) id or is from the input-id which is
mapped to null, depending on whether the id has already been
seen in that particular member of the CVOS. Further alternatives
also arise if a supergroup contains both input-1 and input-2 speaker-ids which
are not mapped.
The overall CVOS is formed from the outputs from all the supergroups. Further simplifications can be applied when generating the final CVOS. For example, giving a single speaker-id to the resegments in supergroups of less than a critical duration can reduce the complexity without detrimentally affecting the output, since the DER is time-weighted and thus supergroups with a small duration seldom impact on the final score. (See  for an example). Alternatively, if the two inputs are very different, one can allow only the labellings corresponding to the two inputs to be passed into the CVOS for a given supergroup thus effectively enforcing an upper-bound on the size of CVOS for each supergroup, hence preventing complexity problems in the subsequent judging stage.
Once the CVOS has been generated a method of choosing the final output must be defined. For the two input case, using confidence scores on the inputs would simply result in the input with the highest confidence being output directly (although this method could be used when there are more than two inputs). Alternative strategies include simple schemes such as assigning the resegments in a supergroup all the same or all different speaker-ids depending on the size of the supergroup, or just picking a member of the CVOS at random, or more mathematically based methods such as using the Bayes Information Criterion (BIC). Here two BIC-based schemes are investigated.
The Bayes Information Criterion gives a log likelihood of the data, L which is penalised in proportion to the number of parameters in the model:
where #M is the number of free parameters, N the number of data points and α the tuning parameter. If K clusters are each modelled using Gaussian(s) of dimension d which have Ni frames and a covariance Si then maximising (1) is the same as minimising:
where P=[0.5d(d+1) + d] for a full covariance Gaussian,
or [2dG+G-1] for a G-mixture diagonal covariance GMM.
For a given member of a supergroup CVOS, a model is made for each output speaker-id. The BMIN value for this set of models given the data segments is then calculated. The final output chosen for the supergroup is the CVOS member which produces the lowest BMIN. The overall output is then simply the concatenation of the final outputs from the (independent) supergroups.
An alternative BIC-based strategy removes the need for the tunable α parameter by ensuring all the model sets being judged contain the same number of free parameters. For example, consider whether to split a parent cluster into two children. A model is built for the children with Mc1 and Mc2 free parameters respectively. The parent model is then built using Mp = Mc1 + Mc2 parameters. The choice between children and parent thus reduces to taking the one that gives the highest likelihood for the data. Here a model is built for each speaker-id as before, but the number of parameters is made proportional to the number of constituent resegments.
The experiments reported in this paper were conducted on the 6-show 2003 evaluation data (bneval03) used in the English Broadcast News RT-03 Rich Transcription evaluations. It consists of 30 minute extracts from 6 different US news shows broadcast in February 2001. Two of these are from radio sources, namely VOA and PRI, whilst four are TV sources, namely NBC, ABC, MNB and CNN. (see  for more details) The diarisation references were generated using the rules described in  using forced alignments provided by the LDC and with 0.3s of silence smoothing applied, and no collars were used during scoring.
The two best diarisation systems from Cambridge University in December 2003
were used as inputs to the cluster voting. They both use the CUED RT-03s
BN segmentation which is based on a GMM speech/music classifier, phone
recogniser and smooth/clusterer, followed by a top-down clustering stage
using PLP coefficients and the arithmetic harmonic sphericity distance
measure. Input-1 uses a BIC-based stopping criterion, whilst Input-2 uses
one based on node-cost. Further details can be found
The DERs on the bneval03 data were 25.12% and 27.09% respectively
although Input-2 was the better system for 5 out of the 6 shows. The
DER is 24.16%(28.05%) if the best(worst) input is taken independently
for each show.
Since the systems were relatively similar, it was possible to exhaustively
search all the combinations of the CVOS to find the best and worst possible
choice for each show [Although the speaker-ids of the supergroups are independent
for both the inputs and the output of the cluster voting scheme, there
is no guarantee that this is also the case for the true reference speakers,
so when scoring against the reference, the supergroups may no longer be
treated independently thus dramatically increasing the complexity -
from 266 possibilities to 24,992 in Experiment 1.
which leads to a DER of 22.79% and 29.44% respectively. These results are
given in Table 1 along with a summary of those from
using the BIC-based judging schemes. More comprehensive results can be found
The results show that the standard BIC technique can be used to reduce the DER to 23.76%, a 1.36% absolute improvement over the best input, whilst the equal-parameter BIC technique can reduce the DER to 23.48%, a 1.64% absolute reduction. Further experiments reported in  show there is a reasonable range in both α value (where applicable) and parameterisation that give an improvement over both inputs.
|Final Output after Judging|
The experiment was repeated using the two-best MIT-LL systems of February
2004. [Thanks to Doug Reynolds for providing the MIT-LL system outputs.
] Input-1 is identical to the system described
in  except that a single full-covariance Gaussian
is used in the agglomerative clustering stage. Input-2 used the same
segmentation and speech/non-speech detection stages, but the clustering
used a system where speakers are represented by their distance to
a set of proxy models. These models were created by
adapting a GMM trained on the entire audio file to each speech segment
in turn. The segments are then represented by a vector of normalised scores
against the proxy models, and the final clustering uses a Euclidean
distance and BIC-style stopping criterion.
The inputs scored 21.38% and 20.59% respectively, but the standards were
considerably different for the individual shows. The scores
from taking the best (worst) input per show separately gave 18.38% (23.59%).
The inputs were very different in places, one supergroup having 39
resegments and thus potentially
members of its CVOS,
so supergroups of more than 12 resegments (generally ~ 500 CVOS members)
simply passed the two input possibilities directly to the final CVOS to
complexity. Similarly, for complexity reasons, the best (worst) possible
scores in the CVOS have been replaced by the best (worst) show scores
seen in all inhouse experiments on the CVOS, which gave a
DER of 16.0% (25.56%).
The results, given in Table 2, again show that
all the experiments give a lower DER than either input except for the
equal-parameter (EP) BIC technique using a single diagonal covariance
The best DER for the EP BIC scheme was 20.33%, a 0.26%
absolute improvement over the best input; whilst the best DER using the
standard BIC scheme was 18.03%, a 2.56% absolute reduction
over the best input.
|Final Output after Judging|
This paper has presented a cluster-voting scheme designed to reduce the diarisation error rate (DER) by combining information from two different diarisation systems. Results on the RT-03 BN evaluation data show the DER can be reduced by 1.64% and 2.56% absolute over the best input when combining the best two systems from Cambridge University and the best two systems from MIT Lincoln Laboratory respectively.
This work was supported by DARPA grant MDA972-02-1-0013. The paper
does not necessarily reflect the position or the policy of the US
Government and no official endorsement should be inferred.