Scientific Report of the 2nd VariAMU Workshop

Following the 1st Variamu workshop at the Hong Kong Polytechnic University in March 2014, the 2nd Variamu Workshop was held on Wednesday 20th and Tuesday 21st August 2014 at the Speech and Language Laboratory, in Aix-en-Provence.

Participants presented their new projects, new collaborations and advancements in on-going collaborations. Besides sessions dedicated to each axis, the working sessions gave room to transversal discussions about variations and corpora, as well as the preparation of the Voice Training 2015, to be held in Hong Kong. The workshop was also the chance to present the experimental platform of the Lab.

The main outcomes of the workshop are presented below.

The detailed programme can be found here.



The core of the language technologies session was based on the integration of variation phenomena in automatic treatment of speech. Different sorts of variation were discussed, from meaning variation to cross-linguistic variations, and variations in pathological speech, shedding light on new methodological proposals (multimodality, ontology and semantics) and in-development tools (SPPAS, MarsaTag).

First, Laurent Prévot emphasised the need for multimodal corpora and analysis to take account of variation at an individual, situational and linguistic level. Through a study of feed-back in several speech situations (dialogal corpora, spontaneous speech, task-orientated speech), his results showed that the most important factor of variation are the meta-data (that is, the speakers). Multimodality opens therefore path to the possibility to compare inter-individual, inter-situational, cross-linguistic and cross-cultural situation.

Brigitte Bigi and Roxana Fung presented a new Variamu project consisting into the development of the software SPPAS into Cantonese. This collaboration emerged to respond to the lack of automatic tool in Cantonese. SPPAS is an automatic speech segmentation and annotation tool based on language-independant algorithms and allowing tokenization, phonetization, alignement, syllabification, repetition, etc. In order to develop the Cantonese version of SPPAS, a pronunciation dictionary, a word-based dictionary and an acoustic model are needed. The Hong Kong Corpus for Adult Cantonese (HKCAC), an 8 hours corpus including detailed phonetic descriptions will provide material for the software. The discussion brought the issues of code-switching treatment, a frequent phenomenon in Hong Kong Cantonese (Cantonese, English, Mandarin).

Quantitative approaches for studying variations was the topic of Emilien Gorenne’s talk. He presented a study of automatic detection of variations in Schizophrenian patients’ speech. 18 features were measured, such as fluency, lexical richness, number of verbs, etc. The outcomes show inter-individual and inter-situational variations.

Grégoire de Montcheuil provided an overview of Ortolang, a collaborative project involving 6 laboratory partners. Ortolang includes a corpus and data repository (SLDR), speech annotation tools (SPPAS and Phonedit), and a speech transcription tool (MarsaTag). The presentation focused then on MarsaTag, a tool combining a tokenizer, a morphosyntax tagger, a POS tagger and a parser. Further developments will consist in building a webservice and a common XML data exchange. So far, the software is developed for written or spoken French input, while the adaptation for Chinese would be possible

Demonstrating that metaphors requires conceptual mapping from source domain to target domain, Chu-Ren Huang proposed to model the conceptual mapping of metaphor on a theory of ontology. Such a conceptual mapping principle offers three methodological and theoretical advantages: 1) to predict possible new uses of metaphors, 2) to provide a classification for non-literal uses, and 3) to measure the conceptual difference of the terms compared.


Working Session 1: CORPORA & TOOLS

The crucial goal of Variamu is to have comparable datasets in order to make cross-linguistic comparisons. For this purpose, a data repository will be created. The repository seeks to host the broadest range of types of speech and metadata: interactional data, second language data, treebanks, aligned corpora for Natural Language Processing, etc.

The repository will allow:

  • to know what can be described;
  • to list what comparable corpora are already done;
  • to know what corpora need to be built;
  • to share corpora and methodologies;
  • to publish a guidelines/best practices for the corpora to be built.

A list of useful corpora and comparable data with brief description will be designed, with the view to create a Variamu data collection on the Speech and Language Data Repository, (part of the Ortolang project). The status of the data (private or public) is still to discuss.

Brigitte Bigi (AMU), Philippe Blache (AMU), Roxana Fung (PolyU), Mantak Leung (PolyU), Grégoire de Montcheuil (AMU), and Stéphane Rauzy (AMU) will take part in the repository project.



The second session gathered works, mainly dedicated to variations both within and across languages. Prosodic, phonetic and syntactical features of Cantonese, English and French were scrutinized, mixing contrastive approaches and structural variations descriptions.

Roxana Fung exposed a comparative study of prosody in two varieties of Cantonese: Hong Kong Cantonese and Guangzhou Cantonese. Measures on vowel space (formants), voice quality, key, register, and speech rhythm were made. First results displays that voice quality and speech prosody are two promising areas to explore phonetic distinction. The outcomes were enlightened by a presentation of evaluative epi-discourses about this variation.

Reporting on the Vowel Variation in Conversational Speech project, Christine Meunier presented the first results. The project studies cross-linguistic variation of vowel reduction in spontaneous speech. To understand this phenomenon, the pronunciation of phonological neighbours is observed in different languages. Outcomes in French shows that, differently than in Spanish and English, higher is the neighbourhood, more speakers hyperarticulate, while perception is more difficult. The next step will consist in a comparison of French results with Cantonese and will include tonal factors, lexical factors (word frequency, etc.), syntactical factors (function words vs. content words) and syllabic factors (monosyllabic vs. plurisyllabic words), as well as speakers’ characteristics (vowel reduction).

If segmental features are frequently under scrutiny in pronunciation studies, supra-segmental features, such as stress, are often neglected. Word stress in English is highly unpredictable and there is a high variation across Englishes. For Hong Kong English, stress may be influenced by Cantonese tones, Cantonese stress and metric rules. After a solid review of the literature, Cathy Wong presented an experimental design methodology to study stress in Hon Kong English. The study seeks to determine if there are HKE stress patterns, and if such, whether they are structural or normative. The results could open path to comparison with other Englishes

Julie Abbou discussed the embedment of gender in language, focusing on the variation of the gender grammaticalization across languages. Through an exploratory study of translation from English to Cantonese, she aimed to compare the linguistic levels and “spaces” of gender in English and Cantonese, two so-called genderless languages. The presentation rose also questions about the status and normalisation of written Cantonese.



This session was dedicated to writing studies and pathology studies.

Chinese writing system components are made up of stroke, logographemes (cluster of strokes), radicals and characters. In his presentation Dustin Lau questioned how to delimit logographeme unit. Based on the hypothesis that the writing duration will be higher between the end of a logographeme and the beginning of the next one, than between 2 strokes from the same logographeme, Dustin Lau made a methodological proposal, consisting in measuring the time distance between stroke writing. The results show that the longest duration lies at the radical boundary. These first outcomes give avenues for next studies, such as the nature of the stimuli (dictation, radical copying, …) or developemental data, such as kid’s sequence of development.

Serge Pinto gave an update of the Speech in bilingual Parkinsonian’s disease patients project, involving Ana Nunes and him. After a presentation of Parkinsonian’s disease symptoms concerning language (among others, speech impairment due to brain damage), Serge Pinto presented the FraLusoPark protocole, gathering France, Brazil and Macau teams with the view to share it with the Hong Kong team.



The Voice Training session is a new collaborative Variamu project.

Gathering AMU, PolyU and Macau U colleagues, this training session will consist in a week of lectures and workshops addressed to the PolyU Speech Therapy School students. The Training will be held in Hong Kong, early March 2015.


Camille Galant (AMU), Alain Ghio (AMU), Antoine Giovanni (AMU), Ana Nunes (Macau U), Serge Pinto (AMU) and Joana Revis (AMU) will join the new Speech Therapy School in Hong Kong for a one-week training in collaboration with Angel Chan (PolyU), Dustin Lau (PolyU), Mantak Leung (PolyU), and Valérie Pereira (PolyU).

Dustin Lau will supervise the event.


The training will consist of:

  • 6 lectures (3h/lecture, 30 students, 2 lectures a days)
  • Rehabilitation practices workshop (1 day, 20 students groups)

The total teaching load will be 18 to 24 hours.

The Aix team will propose a provisional programme and title proposals, based on the following topics:

  • Evaluation
  • Best Practices
  • Evidence in terms of treatment approach
  • Voice and Quality of Life
  • After surgery rehabilitation.

The schedule needs to be checked in accordance with PolyU timetable in the next few days.



Too often, linguistics analyses language by separating the different factors, in order to face the linguistic complexity. However, it is now well evidenced that this separation leads to the loss of crucial elements, since understanding language implies understanding variation.

In addition, variation has long been treated as noise. It is only recently that attempts are made to integrate variation in the analysis and to understand its role in the language complexity. It remains that variation is an imprecise word.

This session shed light on the different dimensions of variation at stake in language studies.

  • The border between variation and language deficiency: is it possible to draw a scale of variation, within pathological speech is integrated?
  • Therapy as be adaptation to other constraints / range of variations
  • Multimodal variation (and inter-modal compensation)
  • Variability of expectations in communication (in terms of representation)
  • Variation as an adaptation to a context
  • Variation as a compensation
  • Language change as a variation and variation in Language change
  • Language evolution along life (perception / production)

Two transversal dimensions completed these topics:

– Variation in prototypical situations vs. Variation in actual productions

– Macro-variations (across languages: different grammar rules,) and micro variation (within a linguistic system: different relations of a given phoneme / word, etc.)

Methodological implications follow this exploration of variation dimensions:

– To measure variation (according to a given state)

– To adapt tools and theories, by taking account of variation

– To use variation measures for linguistic analysis.