The Sheffield language recognition system in NIST LRE 2015

Ng, Raymond W. M.; Nicolao, Mauro; Saz, Óscar; Hasan, Madina; Chettri, Bhusan; Doulaty, Mortaza; Lee, Tan; Hain, Thomas

doi:10.21437/odyssey.2016-26

Cited by 3 publications

(5 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This study proposes to improve the poor performance of phonotactic LR systems, thus tests were focused on 3second and 10-second data only. The language recognition systems were tested on an internal evaluation data set (HELDOUT) constructed by extracting 10% from LDC2015E87 and LDC2015E88 [23], as well as the official LRE2015 EVAL data.…”

Section: Datamentioning

confidence: 99%

“…This study focused on language recognition on speech with 3-second and 10-second durations. In order to remove as much non-speech as possible DNN-based voice activity detection was applied on the raw training data, with the aim to derive speech segments of compatible durations [23]. Segments of required durations in a particular language are then used for sequence training with state-level MBR criterion (Eq(2)) to derive an adapted tokeniser.…”

Section: Unsupervised Adaptation Of Tokenisersmentioning

confidence: 99%

“…However, the Gaussian backend only worked for the adapted tokenisers on 3-second test data. This may be due to the empirical choice of training the Gaussian components on mismatched 30-second training data for the case of unadapted tokeniser [23]. The bottom half of Table 3 shows results for logistic regression score calibration, with the LR system score from the unadapted SWB (1-SWB) tokeniser.…”

Section: Quality Of Tokenisersmentioning

confidence: 99%

“…Despite the marginal difference, the performance gain with the ensemble of tokenisers was carried over in the acoustic-phonotactic fusion system. We also tried system fusion of 20-SWB' with (i) the bottleneck-iVector system [23], and (ii) a combination of the iVector and the bottleneck-iVector systems. All acousticphonotactic fusions demonstrated reduction of minDCF .…”

Section: Fusion With Acoustic Systemsmentioning

confidence: 99%

See 3 more Smart Citations

Combining Weak Tokenisers for Phonotactic Language Recognition in a Resource-Constrained Setting

Chettri

Hain

2016

Interspeech 2016

Self Cite

View full text Add to dashboard Cite

In the phonotactic approach for language recognition, a phone tokeniser is normally used to transform the audio signal into acoustic tokens. The language identity of the speech is modelled by the occurrence statistics of the decoded tokens. The performance of this approach depends heavily on the quality of the audio tokeniser. A high-quality tokeniser in matched condition is not always available for a language recognition task. This study investigated into the performance of a phonotactic language recogniser in a resource-constrained setting, following NIST LRE 2015 specification. An ensemble of phone tokenisers was constructed by applying unsupervised sequence training on different target languages followed by a score-based fusion. This method gave 5 − 7% relative performance improvement to baseline system on LRE 2015 eval set. This gain was retained when the ensemble phonotactic system was further fused with an acoustic iVector system.

show abstract

Section: Datamentioning

confidence: 99%

Section: Unsupervised Adaptation Of Tokenisersmentioning

confidence: 99%

Section: Quality Of Tokenisersmentioning

confidence: 99%

Section: Fusion With Acoustic Systemsmentioning

confidence: 99%

See 2 more Smart Citations

Combining Weak Tokenisers for Phonotactic Language Recognition in a Resource-Constrained Setting

Chettri

Hain

2016

Interspeech 2016

Self Cite

View full text Add to dashboard Cite

show abstract

“…The computational complexity is also affected if the LID system consists of several layers of CNNs, RNNs and DNNs as it requires the re-optimization of each network. Such an LID system's performance significantly deteriorates and it is not easily scalable to accommodate new languages [21][22][23]. The question arises how to effectively scale an existing LID system with minimal performance deterioration and effort required.…”

Section: Introductionmentioning

confidence: 99%

Investigating Scalability in Hierarchical Language Identification System

et al. 2017

View full text Add to dashboard Cite

State-of-the-art language identification (LID) systems are not easily scalable to accommodate new languages. Specifically, as the number of target languages grows the error rate of these LID systems increases rapidly. This paper addresses such a challenge by adopting a hierarchical language identification (HLID) framework. We demonstrate the superior scalability of the HLID framework. In particular, HLID only requires the training of relevant nodes in a hierarchical structure instead of re-training the entire tree. Experiments conducted on a dataset that combined languages from the NIST LRE 2007, 2009, 2011 show that as the number of target languages grows from 28 to 42, the performance of a single level (non-hierarchical) system deteriorates by around 11% while that of the hierarchical system only deteriorates by about 3.4% in terms of Cavg. Finally, experiments also suggest that SVM based systems are more scalable than GPLDA based systems.

show abstract