BUT/Phonexia Bottleneck Feature Extractor

Abstract

This paper complements the public release of the BUT/Phonexia bottleneck (BN) feature extractor. Starting with a brief history of Neural Network (NN)-based and BN-based approaches to speech feature extraction, it describes the structure of the released software. It follows by describing the three provided NNs: the first two trained on the US English Fisher corpus with monophone-state and tied-state targets, and the third network trained in a multilingual fashion on 17 Babel languages. The NNs were technically trained to classify acoustic units, however the networks were optimized with respect to the language recognition task, which is the main focus of this paper. Nevertheless, it is worth noting that apart from language recognition, the provided software can be used with any speech-related task. The paper concludes with a comprehensive summary of the results obtained on the NIST 2015 and 2017 Language Recognition Evaluations tasks.

Spoken Language Identification in Unseen Target Domain Using Within-Sample Similarity Loss

“…1) of the input speech sample. This block contains a pre-trained bottleneck feature (BNF) extractor at the front-end to convert the speech utterance into a sequence of 80-dimensional BNFs [14]. This BNF extractor was originally trained for identifying 3096 phone states from 17 languages.…”

Mel-Weighted Single Frequency Filtering Spectrogram for Dialect Identification

“…The second baseline is an i-vector system which is trained using Mel frequency cepstral coefficients (MFCCs) [20], [40]. The third baseline is an i-vector system, which is trained using the BUT/phonexia bottleneck features [41]. The second and third system will be referred to as the i-vector MFCC system and the i-vector BUT-BNF system, respectively.…”

“…In the latter, BNFs are extracted from a multi-lingual phone recognizer neural network. For our experiments, we considered a pretrained phone recognizer from BUT/phonexia [41] which was trained using 17 Babel languages. For both systems, 100-dimensional i-vectors are extracted using 256 Gaussian mixture components and the obtained i-vectors are transformed by a whitening transformation to be used in the dialect prediction [20], [40].…”

Search on speech from spoken queries: the Multi-domain International ALBAYZIN 2018 Query-by-Example Spoken Term Detection Evaluation

“…The feature extraction module performs stacked bottleneck feature (sBNF) computation following the BUT/Phonexia approach [125], both for queries and utterances. To do so, three different neural networks are The operation of BUT/Phonexia sBNF extractors requires an external VAD module providing speech/nonspeech information.…”