English. CorAIt is a non-native speech database for Italian, which is freely accessible online for academic research purposes. It was especially designed to meet the requirements of a larger research project focused on foreign accented Italian speech. The corpus is aimed at providing a uniform collection of speech samples uttered by non-native speakers of Italian. To date, 105 non-native speakers - whose mother tongues are either French, Romanian, Spanish, English, German, or Russian - have been recorded. The corpus includes also a control group made up of 16 Italian speakers. There are almost 8 hours of audio material, both read speech (first and second reading), and spontaneous speech. This paper emphasizes the necessity for this type of database, it describes the steps involved in its construction, and it presents the features of CorAIt.
Italiano. CorAIt è un corpus audio di l’italiano L2 liberamente consultabile online per scopi di ricerca scientifica. Il corpus è parte integrante di un progetto di ricerca che affronta l’accento straniero nella lingua italiana da una prospettiva più ampia. E’ stato ideato e costruito con lo scopo di fornire una raccolta uniforme di materiale audio prodotto da parlanti di italiano L2. Ad oggi sono stati registrati 105 parlanti stranieri di madrelingua: francese, romena, spagnola, inglese, tedesca, e russa. In aggiunta, il corpus è dotato di un gruppo di controllo composto da 16 parlanti italiani. Sono disponibili circa 8 ore di registrazioni, sia di parlato letto (prima e seconda lettura) che di parlato spontaneo. L’articolo evidenzia la necessità di costruire questo tipo di database, e descrive la progettazione e le caratteristiche di CorAIt. 1
It has become clear that accurately designed
speech corpora are of essential importance for the
development of efficient speech technologies.
Investigating how native and foreign-accented
speech differ is a necessary step in non-native
speech recognition
Currently, the number of non-native speech databases seems almost insignificant if compared to corpora of native speech.
Moreover, until recently, the majority of the
research has focused on English. Therefore, some
of the largest non-native speech databases are
available for this language: TED
Only in the last few years has there been an
interest in other languages. Without claiming to be
exhaustive, some of the largest non-native speech
databases for languages other than English will be
mentioned: BAS Strange I+II
However, as a result of that fact that many
nonnative speech databases are built for commercial
purposes within private research centres, it is
actually quite difficult to map all the resources of
this type ever built
This paper presents CorAIt, a non-native speech database for Italian. The database is part of a Ph.D. project which intends to study foreign accented Italian speech both from a computational perspective (automatic identification and classification of non-native accent) and a perceptual perspective (interpretation of quantitative and qualitative judgments delivered by expert and naïve native Italian speakers with respect to non-native pronunciations). The design and the development of this corpus were determined by several factors, which are outlined below. 2
Currently, the automatic speech recognition systems for Italian which are integrated into generally available virtual assistant software (e.g. Google Now, Google Assistant, Siri, Cortana, etc.) perform quite well on native speech. However, despite recent advances in this field, non-native accents still represent a challenge. This may be due to fact that there is significantly less training data available for automatic speech recognition systems on non-native pronunciations. Considering that Italy is a multicultural country, with over 5 million foreign citizens, representing 8.3% of the entire population residing on its territory1, it would be desirable to provide services to users who speak Italian with non-native accents.
Apart from acting like training sets for automatic speech recognition systems or for textto-speech systems, non-native speech databases might be beneficial in the fields of computer assisted language learning (CALL) and mobileassisted language learning (MALL), as well as for linguistic profiling tasks. Glottologists and scholars working on Italian as a foreign language might also benefit from the presence of these resources.
At the date this research began, there was only
one audio corpus for foreign-accented Italian
speech, freely available for online consultation,
1 The data were provided by the National demographic
balance (year 2016) produced by the Italian National Institute
for Statistics (ISTAT). The full report is available at:
http://www.istat.it/it/files/2017/06/bilanciodemografico2016_13giugno2017.pdf
namely DILS - Dialoghi in Italiano Lingua
Straniera
It is worthwhile to mention that there are
several other learner corpora for Italian: VALICO
- Varietà Apprendimento Lingua Italiana Corpus
Online
In addition to the above-mentioned corpora,
there exists a database of written and spoken
nonnative Italian, entitled ADIL2 - Archivio Digitale
di Italiano L2
The corpus was designed, collected and developed from January 2016 through July 2017, and it was aimed at providing a uniform collection of audio material produced by adult non-native speakers of Italian residing in Bologna2.
Initially, the intent was to collect data for 11 different mother tongues (L1s): Maghrebi Arabic, Urdu, Mandarin Chinese, Albanian, Russian, English, German, French, Romanian, Spanish, and Italian (as a control group). The first 10 groups correspond to the L1s spoken by some of the major foreign populations residing in Italy. However, recruiting speakers for all these groups 2 To simplify the data collection process, the author chose to recruit people that studied or worked in Bologna (her city of residence). proved to be a challenging task. This may be result of the fact that participation was entirely voluntary and no material reward was provided to informants. Since it was not possible to recruit enough speakers of Maghrebi Arabic, Urdu, Mandarin Chinese and Albanian, these four groups were abandoned. 3.1
Specific criteria of quality, quantity and diversity were observed, as much as possible, for each L1 when the participants were selected (Cf. section 4.1).
All speakers were recruited locally in Bologna. Most informants were enrolled as regular or exchange students in B.A., M.A. and Ph.D. programmes at the University of Bologna and they were contacted on their personal e-mail address. The e-mail message contained a description of the research project and informed potential participants about the tasks they would have performed. Nearly one fourth of them replied positively to the call. 3.2
All informants were aware that they were recorded. They gave informed consent in writing to the use of their speech samples and their sociolinguistic data for research purposes.
In order to guarantee uniformity, the same
experimental protocol was employed for all
subjects. Before each recording session, speakers
were asked to fill in a detailed form regarding
their sociocultural and sociolinguistic background.
The digital recordings were performed with a
Samson METEOR MIC cardioid
pickup microphone (condenser diaphragms: 25
mm) on the Praat software
Each recording session lasted around 60 minutes. The sessions were individual-based, and they were guided and monitored by the author. 3.3
The speakers were asked to perform two tasks:
reading an article excerpt published on the Italian
newspaper Corriere della Sera3; and describing
spontaneously how they spent their last holidays.
3 The newspaper article is available online at:
http://cinquantamila.corriere.it/storyTellerArticolo.php?storyI
d=0000002228555. The excerpt was included in this frame:
“Don Geretti è un grande affabulatore […] Pietro terminò il
suo cammino terreno e quello, tormentatissimo, verso la
fede.”
That specific reading fragment was chosen
because it presented various levels of complexity
and it contained all Italian phonemes. The reading
task was necessary for triggering difficulties that
could emerge as a result of conflicting
orthographic conventions between the speakers’
mother tongues and Italian
All participants had two reading attempts and they were asked to read and speak as naturally as they could. 4
CorAIt is a non-native speech database for Italian, which has become fully and freely available online for academic research consultation4. It contains 2,244 audio samples produced by 105 non-native speakers of Italian. It also includes 300 audio samples obtained from 16 native Italian speakers. In total there are almost 8 hours of speech, consisting in roughly 72,000 words. 4.1
Originally, it was planned to recruit at least 15 speakers for each L1. This threshold was reached for French, and it was exceeded for the other six groups.
Regarding the age distribution of the informants, the range is 19-40 years, but most speakers are older than 20 and younger than 30 years (Cf. Table 1).
tongue Russian
Number of speakers 17 16 17 15 20 20 16
Age means
(±S.D.) 27.06 (5.94) 23.75 (4.69) 23.53 (4.06) 23.33 (2.72) 25.40 (2.22) 23.65 (2.95) 29.00 (4.62)
Despite the efforts to call up the same number
of male and female speakers, the corpus is not
perfectly gender-balanced. Apparently, it was
4 CorAIt database is accessible for consultation (prior to
registration) at: www.corait.it
easier to engage female participants. In fact, 80
females (66%) and 41 males (34%) were recorded
for this project. Nonetheless, according to
literature, gender has not been reported as a major
source of pronunciation issues, and for this reason
it was assumed that gender imbalance had minor
effects on this type of studies
For the corpus design, the age of Italian language onset was taken into consideration, and it was distributed as follows: childhood (23%), adolescence (26%), and adulthood (51%). The data are predictable, considering that most informants learnt Italian naturalistically (64%) soon after they moved to Italy. In fact, only 36% of them claimed that they used mainly scholastic methods for learning Italian, and that they had already spoken the language at their arrival in Italy.
Since most informants were exchange students, 59% of them had spent 6-12 months in Italy at the time they were recruited for this research. The remaining part had lived in Italy for 12-24 months (15%), or for more than 24 months (26%). Not surprisingly, the great majority of speakers claimed that they had been exposed only to the Bolognese variety of Italian.
Because it was almost impossible to predict the speakers’ proficiency level5 in Italian before meeting them, the balancedness is not guaranteed for all accent groups (e.g. no Romanian speaker had A2 waystage/elementary level in Italian). For the sake of brevity, at the general level, this variable is represented as follows in the database: waystage/elementary level - A2 (12%), threshold/intermediate level - B1 (28%), vantage/upper-intermediate level - B2 (28%), and advanced/proficiency levels - C1 and C2 (32%). 4.2
An average of 4 minutes and 25 seconds of raw audio material consisting of read and spontaneous speech were recorded for each speaker. Some speakers had to terminate the registration session earlier than planned, so in those cases it was possible to record only their first reading attempt. Regardless of that, the spontaneous speech collected (23%) is, however, inferior to the reading speech material (77%). All raw samples were segmented manually into utterances corresponding to grammatical sentences for the reading material, and to phonological sentences 5 All participants self-assessed their Italian level based on the Common European Framework of Reference for Languages, available at: http://www.coe.int/t/dg4/linguistic/Source/Framework_EN.p df for the spontaneous speech. In most cases, the material was not qualitatively altered, so hesitation phenomena and disfluencies were generally left as they were. 4.3
One contribution of this project is that of making the corpus available to the research community. Following the model of similar tools, a website that would host the database was created. Then, the embedded webapp could extrapolate and classify the audio files from the dataset, according to specific criteria.
For the creation of the webapp, the web framework Django as well as several Python libraries (MySQL-python, django-treebeard, django-filer, html5lib, sorl, wsgi, polymorphic, classy-tags, audiofield, appconf, etc.) were employed. That allowed the use of a powerful ORM system, equipped with a web interface for storing multiple data types into our MySQL database.
Moreover, the Django web framework ORM favoured the realization of data collection models: a model is the only final data source containing the fields and the essential behaviours of the dataset and of the reference objects.
Generally, each model is mapped to a single database table and each attribute represents a database field. The queries are performed by means of ad-hoc APIs for each model.
The project is hosted on a server with a CentOS 7 operating system. CortAIt is already configured for various types of SQL and NoSQL databases (PostgreSQL, MongoDB, Cassandra, etc.). It also supports the execution of some cloud computing platforms, such as Amazon Web Services (AWS), which could improve its performance in case of an exponential growth of the computational complexity. 4.4
The web database is queryable from the dedicated
section of CorAIt website, prior to registration and
approval. Due to storage issues, and observing the
design of Speech Accent Archive
To enable advanced queries, various layers of metadata were added to each audio file: the speaker’s mother tongue, gender, age of Italian language onset, age at the time the sample was recorded, level of Italian proficiency, Italian learning method, length of residence in Italy, proficiency in other foreign languages. Moreover, information on the type of sample and its quality was included (Cf. Figure 1).
The corpus has not been transcribed nor
annotated yet. However, following the example of
the Speech Accent Archive
Besides the embedded audio player – which allows to listen and download the audio sample – the window where the single result is displayed provides biographical and quantitative information with respect to the speaker who uttered that speech sample, as well as qualitative information regarding the audio file (Cf. Figure 2).
Considering that currently this non-native speech database presents some imbalance issues as regards the speakers’ age of Italian language onset, their proficiency level, as well as the length of their residence in Italy, the data collection will be further extended.
In the future, the web database might also be enhanced with orthographic and phonetic transcriptions. Disfluencies (i.e. false starts, filled and silent pauses, phoneme lengthening, mispronounced words), mouth clicks, and external noise could be annotated. 6
The author gratefully acknowledges the informants who participated in this research, and A. M. Tenace, whose expertise and assistance were fundamental for the architecture and the implementation of the webapp.