SPEECHDAT-CAR. A Large Speech Database for Automotive Environments

The aims of the SpeechDat-Car project are to develop a set of speech databases to support training and testing of multilingual speech recognition applications in the car environment. As a result, a total of ten (10) equivalent and similar resources will be created. The 10 languages are Danish, each language 600 sessions will be recorded (from at least 300 speakers) in seven characteristic environments (low speed, high speed with audio equipment on, etc.). This paper gives an overview of the project with a focus on the production phases (recording platforms, speaker recruitment, annotation and distribution). Automatic speech recognition (ASR) appears to be a particularly well adapted technology for providing voice-based interfaces (based on hands-free mode) that will enable new in-car applications to develop while taking care of safety aspects. However, the car environment is known to be particularly noisy (street noise, car engine noise, vibration noises, bubble noise, etc...). To obtain an optimal performance for speech recognition, it is necessary to train the system on large corpora of speech data recorded in context (i.e. directly in the car). For this reason, language-specific initiatives for database collections have been developed since about 1990 [Langmann (1998)]. The European project SpeechDat-Car 1 aims at providing a set of uniform, coherent databases for nine European languages and for American English. in developing large-scale speech resources for a wide range of languages and for in-car applications participation of external partners to the original consortium is also possible. Siemens is an 'external' partner. It is also important to note that SpeechDat-Car commits itself to a strict validation protocol to ensure 1 SpeechDat-Car started in April 1998 in the 4th EC framework under project code LE4-8334 with a 30 months' project duration. optimal quality and exchangeability of the databases [Van den Heuvel (1999)]. This paper gives an overview of the project with a focus on production phases. It is organised as follows: the next section describes the database specifications (database content, recording platforms and validation procedures). Then, Section 3 provides additional information on speaker recruitment and an extensive description of the annotation procedure and tools is given in section 4. The paper then concludes with a short section about database availability and dissemination. Each database produced in the SpeechDat-Car project is intended to provide enough data to adapt speaker independent recognition systems to the automotive environment. Database contents were designed to cope with different applications. The design …