This paper investigates the use of deep neural networks (DNNs) for the task of spoken language identification. Various feed-forward fully connected, convolutional and recurrent DNN architectures are adopted and compared against a baseline i-vector based system. Moreover, DNNs are also utilized for extraction of bottleneck features from the input signal. The dataset used for experimental evaluation contains utterances belonging to languages that are all related to each other and sometimes hard to distinguish even for human listeners: it is compiled from recordings of the 11 most widespread Slavic languages. We also released this Slavic dataset to the general public, because a similar collection is not publicly available through any other source. The best results were yielded by a bidirectional recurrent DNN with gated recurrent units that was fed by bottleneck features. In this case, the baseline ER was reduced from 4.2% to 1.2% and Cavg from 2.3% to 0.6%.

Slavic languages pose several specific challenges that need to be addressed in an ASR system design. Since we have already built an engine suited for highly-inflected languages, we focus on adopting it for new languages, now. In this case, we present an efficient way to adapt the system to all (seven) South Slavic languages, using methods and tools that benefit from language similarities, easily adjustable G2P rules or common phonetic subsets. We show that it is possible to build accurate language and acoustic models in an almost automated way, entirely from resources found on the web. The AMs are trained via cross-lingual bootstrapping followed by lightly supervised retraining from public data, like broadcast and parliament archives. Tests done on a set of main broadcast news in each language show WER values in range 16.8 to 21.5 %, which includes also errors caused by OOL (out-of-language) utterances often occurring in this type of spoken programs.

No abstract

No abstract

No abstract