General conclusions

• Among existing and prospective resources, usability and reusability of speech databases produced through EEC projects should be enhanced. These resources are most of the time still unavailable several years after the end of the projects, because the time-intensive process leading from the raw data to marketable data was not initially included (or funded) in the projects (the marketable product is not a deliverable of the project). The distribution of these databases can be stimulated by the existence of an European Center for the Distribution of Language Resources who can provide funding for reorganisation and documentation of the already existing data and its production on CD-ROM.
• Existing databases whose diffusion is confidential should be highlighted. These databases are either not well-known because they remain in the scope of the project for which they have been designed, and/or have been kept confidential by their designers who do not want to disseminate them. It is obvious that for many industrials, linguistic resources are not yet considered as precompetitive resources, but as strategic ones. Any future European Center of LR production and distribution will have to manage the gap between the views and methods of industrials and academics: long-term widely available resources for academics vs. short-term strategic resources for industrials.
• There is a clear lack of speech databases suitable for multilingual evaluation. At the curent time, only EUROM has a Europe-wide dimension. Large, multilingual corpora are needed, and added value (full annotation and labelling) must be included in the design of new databases. It should be noted that the most widely used multilingual corpora are currently being recorded by OGI in the U.S. This corpus is being used for development and testing of automatic language identification techniques, which have a wide variety of practical applications.
• Pronouncing dictionaries and spoken language lexicons are very limited and should be developed for speech output systems assessment. Large phonetic lexicons are needed for use in speech recognition. One particular source that should be exploited is the outcome of the Onomastic project, and in particular the legal details for distribution of the lexica need to be worked out.
• A corpus of at least 100 hours of high quality speech should exist in each language providing coverage of many words and phonetic contexts, with data from a reasonably large number of speakers (200-500). Most languages are far from that (for example at 20kHz, 100 hours of speech correspond at 15 Gbytes, or about 25 to 30 CD-ROMs). In addition, this basic corpus will need to be supplemented by application-oriented corpora (such as for dictation (medical, legal, insurance domains), office systems, speaker verification/identification, topic spotting, information retrieval, etc.). Newspaper-based corpora are a potential multilingual source that should be encouraged, as these basic corpora provide at the same time speech and language modeling data, that can be used for dictation as well as other applications such as topic spotting. Multi-sensor and multimodal corpora are also needed for more basic research.
• For short-term commercial use there is a need for large telephone-based corpora, with speech from many (several thousands of) speakers covering a wide range of dialect, age, and socioeconomic backgrounds. These corpora should be extensive enough to permit the design of speaker-independent, vocabulary-independent speech recognisers that can serve as the source for a variety of applications.

Our conclusions are in agreement with those in the recent EAGLES report on Spoken Language Systems. It is clear that the need for adequate resources is a prime concern for many actors in the field of language technology. The following excerpts come from a draft EAGLES report on Spoken Language Systems (For EAGLES Restricted use). We highlight three areas of needs: speech corpora, lexicons and assessment of speech output.

• Speech corpora size (in terms of speakers), pp. 10-11
  
  • few speakers (< 5 speakers):for development of speech synthesis systems (dictionaries of phonetic elements). Advanced research. Multi-channel recordings (Electroglottogram, subglottal pressure, etc.)
  • medium (5-50 speakers): experimental research. Number of speakers and repetitions large enough for statistical processing / or for broad coverage of phenomena.
  • large (> 50 speakers): Train and test of speaker-independent recognition systems.
• Spoken language lexicons (p. 67)
  ``Large-scale spoken language lexical resources, from reference sources of standard, stylistic and regional pronunciations through vocabularies which are characteristic of spoken language, are required for current research and development with both statistical and knowledge-based technologies ...
  These spoken language lexical resources in the form of actual lexical databases and tools for constructing them, are sadly lacking.''
• Assessment of speech output systems (p. 102)
  ``A short-term recommendation is to develop multilingual machine readable pronouncing dictionaries at the single word level which list permissible variations...''