The problem of labelling speaker turns by automatically segmenting and clustering a continuous audio stream is addressed. A new clustering scheme is presented and evaluated using a clustering efficiency score which treats both agglomerative and divisive clustering strategies equally. Results show an efficiency of 70% can be obtained on both manually and automatically derived segments on the 1996 Hub4 development data.
For the task of identifying potentially unknown anchor speakers within broadcast news shows, the frame classification error rate is very important. To reflect this, a frame-based cluster efficiency is defined and the results show a 90% frame-based efficiency can be achieved. Finally a frame-based comparison between the manually and automatically derived segment/cluster sets shows that approximately one third of the errors are introduced during segmentation and two-thirds during clustering.
In recent work we have described automatic methods for both segmenting and clustering a continuous audio stream input [2, 4] . These methods were shown to be an important part of our overall recognition system for broadcast news. The segmenter is designed to produce segments of of between 1 and 30 seconds duration which are acoustically homogeneous (i.e. they contain only one speaker and noise/channel condition). The clusterer is designed to place acoustically similar segments into groups (clusters) of a certain minimum occupancy (generally 30 seconds). This allows Maximum Likelihood Linear Regression (MLLR) to be applied, thus improving the overall performance of the recognition system.
In this work, attention is switched to the problem of determining speaker turns when the speakers (and number of speakers) are unknown. The aim therefore is to produce just 1 pure cluster for every speaker, independent of the amount of time the speaker is talking. The previous speaker-adaptation clustering strategy is modified and a new recombination procedure introduced, to reflect this new aim of speaker-identification.
Initially the clustering performance for both these systems is evaluated on an utterance basis. This reflects the task when the user has a database of recorded utterances and wishes to retrieve one from a given speaker as quickly as possible. By presenting perfect speaker clusters the number of utterances the user has to listen to in order to find the appropriate message is dramatically reduced. The clustering efficiency from  is used to present the results and it is shown that setting the free parameter can produce results which equate the no-clustering case in both divisive and agglomerative clustering.
The performance measure is then moved from an utterance-level to a frame-level basis. This allows greater emphasis to be placed on longer segments and models the task of tracking (unknown) speakers through a broadcast news show when the user may not be interested in very short utterances. The frame-based approach also allows the separate errors introduced by the segmentation and clustering stages to be quantified.
This paper describes briefly the segmenter and clusterer in section 2, introduces the clustering performance measures and derives formulae for the critical case in section 3, gives experimental details in section 4, presents utterance-based and frame-based results in sections 5 and 6 and offers conclusions in section 7.
The clusterer (described in ) represents each segment by a single correlation matrix. The arithmetic harmonic sphericity  is used as the distance measure. A top-down split-and-merge algorithm is used for the clustering. Each node is split into 4 child nodes and the new correlation matrices for the child nodes are calculated by concatenating the data within them. The segments are then assigned to the closest child node, the statistics recalculated and the process repeated until no more segments move. This is repeated until all the nodes have been split completely.
It is necessary to define when a split is allowable to prevent the data being split back into its constituent segments. The speaker-adaptation scheme sets a minimum occupancy requirement of 3000 frames (30 seconds) on the final clusters to ensure robust speaker adaptation can follow. For the speaker-identification scheme no such restriction is necessary and alternative stopping criterion must be found. New parameters are introduced which model the minimum gain required from splitting and the maximum level of overlap between child nodes to allow the split to go ahead. Another parameter is added to deal with the special case of singleton clusters where the intra-node distance is zero. By changing these parameters whilst keeping the minimum occupancy required to zero, different levels of recombination for the speaker-identification scheme can be achieved.
|Ns||Total number of speakers|
|Nc||Total number of clusters|
|Nu||Total number of utterances|
|nij||# utterances in cluster i from speaker j|
|# utterances said by speaker j|
|# utterances in cluster i|
|purity of cluster i|
The Rand Index
The first metric used in this paper is the Rand Index .
This gives the number of utterance pairs that are from the same speaker and are not in the same cluster or that are from different speakers but are in the same cluster. Smaller IRAND therefore represents a better speaker split, with perfect speaker split having an IRAND of zero.
The second metric used is the clustering efficiency from . This is based on the BBN metric:
where Q is a user-defined parameter which represents the trade off between producing a few large clusters which may contain multiple speakers and incomplete clustering where certain speakers may have more than one cluster associated with them.
Clustering Efficiency is then defined in terms of perfect clustering, I(P), and the singleton cluster set, I(S), which represents the case of no clustering for an agglomerative scheme. Note that this value is not a true efficiency as it is possible to obtain a negative value for .
For the singleton clusters (each utterance is a cluster):
and Nc = Nu so I(S) = Nu(1-Q) .
For perfect clustering:
and Nc = Ns so I(P) = Nu-QN s.
With this metric perfect clustering produces a score of 1.0 whilst the singleton cluster set scores 0.0. Note however, that another limit on performance exists, namely grouping all the utterances into 1 large cluster. This may produce a negative efficiency score, depending on the choice of Q.
Experiments with Q set to 0.5 are reported in this paper to allow comparisons with previous work in this area [6, 5]. However, this gives an efficiency of around -1 for the case of a single cluster for the data used in this paper. It would be nice to have a baseline score of zero for the case of no clustering irrespective of whether the clustering is implemented in a divisive of agglomerative scheme. To achieve this, the value of Q is set to a critical value such that the one-cluster case also has a cluster efficiency, I(1), of zero:
For 1 cluster: Nc = 1; ni = Nu for i=1 and 0 otherwise
hence setting I(1)=I(S) so that gives:
Note that since:
For the experiments reported in this paper, .
Results are presented for the cases of one overall cluster, (one_c), singleton clustering, (singleton_c), perfect clustering, (perfect_c), speaker-adaptation clustering, (adapt_c), and two speaker-identification systems (speak_1_c, speak_2_c). The speaker-adaptation scheme is that used in our overall recognition system before speaker adaptation , whilst the speaker-identification systems use the scheme described in section 2 with different levels of recombination. The automatic segmentation is done using our 1997 segmenter .
Section 5 reports the results using the utterance-based metrics described in section 3 for the cases of Q=0.5 and Q=Qcrit Section 6 uses the same cluster sets but gives the results on frame-based metrics.
It is interesting to note that the results for the critical value of Q show a slightly different pattern, namely that the speaker-adaptation scheme scores higher than the speak_2_c scheme, due to the smaller number of clusters.
These results are very similar to the manually-segmented case and show the same trends, namely that speaker-adaptation clustering gives a reasonable performance in the utterance-clustering task, but the performance can be increased further by switching to the speaker-identification scheme. The approximation that each segment only contains the dominant speaker does not seem to affect the results unduly.
In order to be able to look at the relative effects of automating both the segmentation and the clustering on the overall performance, the definition of the scoring metric must be redefined to work on a frame basis. [As the initial segments are not the same for the manual and automatic case.] This also reflects the true performance on certain tasks more accurately than utterance-based metrics. For example, for the identification of a (potentially unknown) anchor speaker in a broadcast news show, an error with a long utterance may be more significant than an error with a shorter utterance. A new frame-based efficiency is therefore defined. The previous formulae remain the same but the definitions are altered to:
|Nf||Total number of FRAMES|
|nij||# FRAMES in cluster i from speaker j|
|# FRAMES said by speaker j|
|# FRAMES in cluster i|
and the resulting baseline cases become:
|singletons (each frame separate):|
The results for the cluster sets given in section 5 recalculated using this frame-based score are given in Tables 3 and 4 for the manual and automatically derived segments respectively. The frame rate was 100Hz and the number of frames after segmentation was approximately 600,000.
These results show that for manual segmentation a very high frame-based clustering efficiency of 81% (90% for Qcrit ) can be obtained from automatic clustering. For the automatic segmentation compared to the automatic baseline, the results are almost identical to the manual case, with a frame-based efficiency of 79% (89% for Qcrit ).
The score from comparing the frame labels from the automatically segmented/clustered set with the perfectly segmented/clustered set are given in Table 5. Note that the Rand Index for the perfect case and the cluster efficiency at the previous value of Qcrit are no longer zero, due to errors introduced in the segmenter when non-speech events are removed.
These results for cluster efficiency (Q = 0.5) are summarised in Table 6. After the automatic segmentation has occurred, the perfect speaker clustering results in 89.2% efficiency (as compared to the manually generated speaker baseline). Automatic clustering of these segments results in this number falling to 72.3%. Note the drop due to automatic clustering is around 18% for both the manual and automatic segmentation confirming that the errors introduced in segmentation are largely independent of those made in clustering.
These results show our methods of automatic segmentation and clustering produce a frame-based efficiency of 72.3% on the 1996 Hub4 development data. The loss of 27.7% from the perfect case is 39% due to errors in the automatic segmentation process, and 61% from the clustering procedure.
The concept of clustering efficiency has been extended to score divisive and agglomerative clustering schemes evenly, and a new frame-based scheme has been introduced. It is interesting to note that the relative performance of both the speaker-identification clustering schemes and the speaker-adaptation scheme depends on which definition of efficiency is used. In tasks such as following (potentially unknown) speakers through broadcast news shows, where frame error is more important, the speaker-identification system with moderate recombination performs the best. It is clear that to get optimal performance in segregating speakers, the task must be clearly defined before deciding how to run the clusterer.