All our systems are based on Dynamic Time Warping (DTW). These systems use bottle-neck features (BN) as input. The bottle-neck feature extractors were trained on GlobalPhone Czech, Portuguese, Russian and Spanish languages, so our approach is in low-resource category. We also aimed on T1/T2/T3 types of query search for late submission systems. System calibration and fusion were based on binary logistic regression.
We developed one (single) system for on-time submission
and two more systems for late submission. The system
schema is in Figure 1. Similarly to last year, we used
feature extractors already available at BUT (so-called Atomic
Systems). We aimed only at bottle-neck features and DTW
search approach this year. Our goal was to build a simple
system and aim on word reordering in queries (T2/T3) (we
addressed this problem in late submission). On the other
hand, we have not addressed noise and reverberation in the
data (see [
All our subsystems use Arti cial Neural Networks (ANN)
to estimate per-frame phone-state probabilities (so-called
posterior-grams) and bottle-neck features. Subsystems are
based on DTW using BNs to calculate distances between
query and test segment frames. We re-use ANNs, which were
trained for different projects as acoustic models for phone
or LVCSR recognizers: 1 SpeechDat (Hungarian;
monolingual LCRC systems [
We ended up with 4 atomic systems and 3 subsystems based on DTW using GlobalPhone features. Data
sAytostmemic GQP CZ BN GP PO BN GP RU BN GP SP BN
U
Subsystem Normalization
Calibration fea stack norm calib
DTW 2w
DTW 3w norm calib
norm calib late submission Output detections 2.1
We use concatenation of feature vectors for DTW
proposed by GTTS [
In our implementation, we follow the standard
query-byexample recipe { sub-sequence DTW [
In our implementation, we compare the accumulated distances (including the current local distance) normalized by online normalization. The online normalization performs the division by current path length on-the- y for every step p-fea stack DTW l-fea stack DTW+slope l-fea stack DTW 2w+slope l-fea stack DTW 3w+slope l-fea stack DTW+slope+2w3w fusion sideinfo
QU
eval dev calculation to decide which step (vertical, horizontal or diagonal) is the best to choose. The division is not saved during the calculation, it is performed only to decide the next step. The length normalization is done afterwards as in standard approach. This leads to prefer longer paths over shorter ones.
As the distance metric, we used the Pearson product
moment correlation distance [
We used different step size conditions during the calculation of accumulated distances to control the slope of paths (systems denoted as +slope). Each path has a local slope within the bounds 12 and 2. This limitation allows us to eliminate errors where one query frame maps to the whole utterance perfectly or vice versa. We also experimented with different local weights for vertical, horizontal and diagonal direction but we got no improvement out of it. 3.1
We built additional subsystems to deal with T2/T3 type of queries. A query is split into equal parts and for each part, DTW is performed separately (denoted as bands). The smallest accumulated distance is chosen from each band of the given query and results are averaged together as a matching score. This approach allows us to search for multiple word queries. For single word queries, the results remain the same. Note that it is not mandatory that two words in T2 query are separated exactly in the middle of the query. We experimented with 1 (baseline system), 2 (denoted as 2w) and 3 (denoted as 3w) bands. These subsystems were used only for fusion and, therefore, were not submitted as separated late systems.
The global minimum of frame-by-frame detection scores is selected as candidate detection. There might be signi cant differences between the score distributions corresponding to the different queries and it is important to normalize the scores for each query.
We applied m-norm (developed in SWS2013 [
As the task expects only one score per query{utterance pair without timing, we nd and return the best particular score from a set of detections of a query in an utterance. 5.
The post-processed scores were calibrated with respect to the Cnxe scoring metric using binary logistic regression.
We attached a sideinfo to each score (query{utterance pair). The sideinfo consists of: number of phonemes, log of number of phonemes, number of speech frames, log of number of speech frames and average log-posterior of speech frames taken from SAD. The sideinfo was generated for queries and utterances so the nal \feature vector" for calibration consists of: 1 detection score (query{utterance pair), 5 query sideinfo, 5 utterance sideinfo. Parameters (11 linear weights and 1 additive constant) were trained on development set. We denoted this 10 sideinfo parameters as QU. However, we found sideinfo harms the performance for late submission so we omit it.
We applied fusion on the level of calibrated systems using the binary logistic regression again.
For improved late system, we fused the primary system with slope limitation (l-fea stack DTW+slope) and the 2w and 3w systems. The fused system is denoted as \l-fea stack DTW+slope+2w3w fusion" and was submitted as second late system. 7.
First, we processed data by the primary system without respect to T2 and T3 type of query. The rst general late system using slope constraint improved output score by 0:79% in Cnmxien for eval data. The conclusion is that for such noisy data, there could be one or few single query frames tting perfectly to single or few frames from utterance. This generates path too steep or too shallow, obviously not a matching hit.
To improve accuracy of T2 queries, we used algorithm where queries are split into parts and search is performed \per-partes". The output scores of these subsystems were not signi cantly better, however, helped in fusion with the best single system. We got improvement of 0:6% in Cnmxien. More detailed, the deterioration of T1 query is 0:15% but there is slight improvement for T2 query (0:71%) and T3 query (0:94%).
The real-time factor for the primary system is 0:009, for late system without fusion is 0:009 and for late system with fusion reaches 0:023. The highest memory consumption (high level water mark) is 450MB. The experiments were run on a hybrid cluster with average CPU Intel(R) Xeon(R) CPU X5670 @ 3GHz.