Unsupervised speech transcription and alignment based on two complementary ASR systems

TEDxSK and JumpSK: A New Slovak Speech Recognition Dedicated Corpus

“…for initial experiments with unsupervised transcription, annotation and acquisition of large speech databases we have created a new speech recognition system architecture based on the complementarity of two Slovak LVCSR systems (see fig. 1, LVCSR 1 and LVCSR 2) [1], [17]. The Slovak LVCSR system uses an open-source recognition engine Julius [18] that was modified to support multi-threaded parallel speech recognition and sharing acoustic and language models among all instances for memory space saving purposes.…”

“…The first acoustic model (AM 1) was trained on 320 hours of manually annotated speech recordings of judicial readings and parliament proceedings [20]. The second model (AM 2) was trained on a database of 330 hours of manually annotated speech recordings acquired from the main broadcast news [21] and Court TV shows with a high degree of spontaneity [1], [21]. Both acoustic models (AM 1 and AM 2) were generated from feature vectors with standard dimension of 39 mel-frequency cepstral coefficients, along with delta and acceleration coefficients and cepstral mean normalization enabled.…”

“…Typical manual transcription speeds of spontaneous or conversational speech lasts around 7 to 12 times real-time, due to its complexity. The transcription and annotation of non-native speech is an even more difficult, slow and laborious process [1].…”