Speech Processing Group

This shell package allows to extract features based on Region Dependent Transforms (RDT) models from audio files. The features are well suitable mainly for Gaussian Mixture Models (GMM) in Automatic Speech Recognition (ASR) systems but they could be used in other applications as well.

The whole process could be split into 3 steps. Standard PLP-HLDA features are concatenated with Stacked Bottle-Neck Features trained in multilingual fashion on Babel data coming from 17 different languages. This features are going into discriminatively trained RDT transforms on 17 Babel languages which generates final outputs.

This python code implements speaker diarization algorithm described in:
http://www.fit.vutbr.cz/~burget/VB_diarization_slides.pdf

This algorithm is based on a generalized version of the model described in:
Kenny, P. Bayesian Analysis of Speaker Diarization with Eigenvoice Priors, Montreal, CRIM, May 2008, http://www.crim.ca/perso/patrick.kenny/BayesCluster.pdf

Kenny, P., Reynolds, D., and Castaldo, F. Diarization of Telephone Conversations using Factor Analysis IEEE Journal of Selected Topics in Signal Processing, December 2010, http://www.crim.ca/perso/patrick.kenny/Kenny_sdfa.pdf

The generalization introduced in this implementation lies in using an HMM instead of the simple mixture model when modeling generation of segments (or even frames) from speakers. HMM limits the probability of switching between speakers when changing frames, which makes it possible to use the model on frame-by-frame bases without any need to iterate between 1) clustering speech segments and 2) re-segmentation (i.e. as it was done in the paper above).

Download: http://www.fit.vutbr.cz/~burget/VB_diarization.zip (attention, the file has 35 MB as it contains example data).

Link to database: http://speech.fit.vutbr.cz/files/quesst14Database.tgz [1 116 MB]

The QUESST 2014 search dataset consists of 23 hours or around 12.500 spoken documents in the following languages: Albanian, Basque, Czech, non-native English, Romanian and Slovak. The languages were chosen so that relatively little annotated data can be found for them, as would be the case for a ``low resource'' language. The recordings were PCM encoded with 8 KHz sampling rate and 16 bit resolution (down-sampling or re-encoding were done when necessary to homogenize the database). The spoken documents (6.6 seconds long on average) were extracted from longer recordings of different types: read, broadcast, lecture and conversational speech. Besides language and speech type variability, the search dataset also features acoustic environment and channel variability. The distribution of spoken documents per language is shown in Table 1. The database is free for research purposes. Feel free to use the setup for evaluation and comparison of you results with results achieved by others and MediaEval QUESST 2014 evaluations.

Here (http://ceur-ws.org/Vol-1263/mediaeval2014_submission_35.pdf) you can find the MediaEval QUESST 2014 task description, and here (http://ceur-ws.org/Vol-1263/) are the system descriptions of particular teams which participated in MediaEval QUESST 2014 evaluations. An overview paper discussing achieved results is published at ICASSP 2015 and is available here. If you publish any results based on this QUESST2014 database, please refer the paper (bibtex).

Table 1:

According to the spoken language and the recording conditions, the database is organized into 5 language subsets:

Neural network based language models are nowdays among the most successful techniques for statistical language modeling. They can be easily applied in wide range of tasks, including automatic speech recognition and machine translation, and provide significant improvements over classic backoff n-gram models. The 'rnnlm' toolkit can be used to train, evaluate and use such models.

The goal of this toolkit is to speed up research progress in the language modeling field. First, by providing useful implementation that can demonstrate some of the principles. Second, for the empirical experiments when used in speech recognition and other applications. And finally third, by providing a strong state of the art baseline results, to which future research that aims to "beat state of the art techniques" should compare to.