Speech recognition for under-resourced languages represents an active field of research during the past decade. The tunisian arabic dialect has been chosen as a typical example for an under-resourced Arabic dialect. We propose, in this paper, our first steps to build an automatic speech recognition system for Tunisian dialect. Several Acoustic Models have been trained using HMM-GMM and HMM-DNN system. The speech corpus has been collected and transcribed from dialogues in the Tunisian Railway Transport Network. The HMM-DNN system can give an impressive relative reduction in WER.
The remainder of this paper is organized as follows: section 2 introduces Tunisian dialect. Section 3 presents the ASR and the different component of an ASR system. A brief related works for MSA and dialect will be presented in section 4. In section 5, we will present our differents models for building an ASR system for Tunisian dialect and discussing the experimental results. Section 6 concludes this work. 2
Tunisian dialect is the spoken Arabic language in Tunisia. It is used in daily life and
emerged as as the language of communication online: social media, blogs, SMS, emails,
etc. It exists many regional varieties of Tunisian dialect and it differs from a region
to another. This varieties includes the Tunis dialect (Capital), Sahil dialect, Sfax
dialect, Northwestern Tunisian dialect, Southwestern Tunisian dialect, and Southeastern
Tunisian dialect [
Automatic Speech Recognition (ASR) is a tehnologie that converts an audio signal to text. Speech recognition software is now frequently installed in computers and mobile devices, allowing for easy access and makes life easier.
ASR has a wide range of applications like command recognition, dictation, interactive
voice response, learning foreign language [
Researchers on automatic speech recognition have several potential choices of
opensource toolkits for building a recognition system such as HTK [
The waveforme audio given in the input is converted into a sequence of fixed size acoustic vectors Y=y1,...,yn. The decoder tries then to find the sequence of words W=w1,...,wn which is which is most likely to have generated Y.
W^ =argwmax{P(Y|w)P(w)}
P(Y|w) is determined by an acoustic model and P(w) is determined by a language model. 3.1
The first step in any ASR system is to extract features. It consists in identifying the component of an audio signal that are useful for the recognition task and discarding all the remaining useless information such as noise, emotion, etc.
Mel Frequency Ceptral Coefficient (MFCC) is one of the common methods for feature extraction. The signal is framed into 20-40 ms frames. The intuition underlying this assumtion is that on short time scales the audio signal doesn’t change much and we assume that the signal is enough stable. Afterwards, for each frame a vector of acoustic parameter is extracted. 3.2
Acoustic modeling represents the relationship between linguistic units of speech and audio signals. An acoustic model is created by taking a large database of speech and using special training algorithms to create statistical representations for each phoneme in a language. These statistical representations are called Hidden Markov Models (HMM). Each phoneme has its own HMM.
A Hidden Markov Model (HMM) is a statistical model for representing probability distributions over sequences of observations. A HMM consists of two stochastic processes, an invisible process of hidden states and a visible process of observable events. HMMs models are trained on data with the forward-backward or Baum-Welch algorithm. In speech recognition, the speech recognizer tries to find the sequence of phonemes (states) that gave rise to the actual uttered sound (observations).
The HMM relies on the assumption that the probability of being in a state at time t depends only on the state at time t-1. Formally: P(zt|zt 1,zt 2, ...)=P(zt|zt 1)
If the acoustic model is trained with sufficient number of speakers, it will be able to represent the properties of new speakers. This model is called speaker-independent. If the corpus includes the data of a specific speakers, such as a model trained with speakers belonging to the same community, it is considered speaker-dependent. 3.3
The Pronunciation Dictionary (PD) or lexicons has an important role in the predictive powers of ASR. It maps vocabulary words to sequence of phonemes which indicates the pronunciation of each of these words. For example: hello H EH L OW. Lexicon should cover all the words we need, otherwise the system will not be able to recognize them. To fix this problem, we need a langage model (we will present it in the next section); the system looks for the word both in the lexicon and in the language model. 3.4
The Language Model (LM) is used to estimate the probability of a word given the word sequence that has already been observed:
P(wn|w1,w2,w3...wn 1)
We distinguish two types of langage model: Grammar Based Language Models (GBLM) and Probabilistic Language Models (PLM). The latter are currently dominate the field of speech recognition because GBLM is not generally useful for large vocabulary applications and it is so difficult to write a grammar with sufficient coverage. N-gram models are the simplest kind of statistical language model. The basic idea is to consider the structure of a corpus as the probability of different words occurring alone or occurring in sequence. 3-gram are probably the most common ones used in ASR and represent a good balance between complexity and robust estimation. In a 3-gram language model the probability of a word given it’s predecessors is estimated by the probability given the previous two words:
P(wn|w1,w2,w3,w4,...wn-1) = P(wn|wn-2,wn-1) 3.5
Speech decoder is one of the central parts of ASR system. Decoding calculates which sequence of words is most likely to match best with speech given an acoustic and langage models. The decoder listens for the distinct sounds spoken by a user and then looks for a matching HMM in the Acoustic Model. If it is true, it keeps track of the matching phonemes. Then, the decoder looks up the matching series of phonemes it finds in its Pronunciation Dictionary to determine which word is spoken. 4
Related works
[
[
The experiments showed better results when using a normalized text than a raw or normalized and vowelized text for the LM using the SGMM+bMMI AM. The best results are obtained using SGMM+bMMI,DNN and DNN+MPE having respectively 30.73%, 29.81% and 26.95%.
[
Firstable, the Qatari LM was interpolated with the MSA-LM which is trained using the LDC Gigaword corpus. this interpolation resulted in a vocabulary size of 265.7K words. Thereby, a significantly decrease of out of vocablary (OOV) rate. Afterwards, a MSA acoustic model is used to decode qatari speech data. A data pooling technique was used after the previous step consisting on training an acoustic model using both qatari and MSA data.
Then, an acoustic aodel adaptation was applied using Maximum Likelihood Linear Regression (MLLR) and Maximum APosteriori (MAP) re-estimation on the MSA model using Qatari speech Data. A combination of different system was applied finally leading on 21.3% and 28.9% relative WER reduction on QA development set and evaluation set respectively. The MSA corpus consists of two speech resources: the NEMLAR Broadcast News Speech Corpus which consists of about 40 hours of audio, and the NetDC which has about 22.5 hours of Arabic Broadcast News Speech. The Qatari corpus consists of 15 hours collected from different TV series and talk show programs.
We used Kaldi Recognition toolkit [
We used for our experiments the TARIC corpus (Tunisian Arabic Railway Interaction
Corpus) [
Data #hours #vocablary Train 8 Hours and 57 Minutes 3027 Dev 33 Minutes and 40 Seconds 612
Test 43 Minutes and 14 Seconds 1009 Different models are trained using HMM-GMM and HMM-DNN systems during acoustic model training. The first one was trained using MFCC, deltas and deltas-deltas features. Linear Discriminant Analysis (LDA) and Maximum Likelihood Linear Transform (MLLT) are applied to reduce dimensionality, which improves accuracy as well as recognition speed. Speaker Adaptive Training (SAT) was introduced afterwards with Feature-space MLLR (fMLLR).
A Maximum Mutual Information (MMI) criterion is used also to give a high discriminative ability to the system and thus MMI belongs to the so called "discriminative training" category.
Finally, we trained a DNN model on top of fMLLR features. The training is done in three stages: RBM pre-training, frame cross-entropy training which has as objective to classify frames to correct pdfs, and the sequence-training optimizing sMBR. The DNN training was done using one GPU on a single machine.
Our LM was built using TARIC transcription. We have trained a 3-gram model given the training corpus and the vocablary using SRILM toolkit. The given vocablary has 3211 unique words. 5.4
In this paper, we present our work on establishing Kaldi recipes to build Tunisian speech recognition system. Different Acoustic models have been trained using different techniques in order to increase system performances. The best results are coming from the training of DNN models, with an overall WER of 25% for dev set 36.8% for test. As a future work, we will ameliorate this results by increasing our data set, and using crosslingual approach to take benifit from the other langages.