A Real-life, French-accented Corpus of Air Traffic Control Communications

A Real-life, French-accented Corpus of Air Traffic Control Communications Estelle Delpech​1​, Marion Laignelet​2​, Christophe Pimm​2​, Céline Raynal​2​, Michal Trzos​3​, Alexandre Arnold​1​, Dominique Pronto​1 1​AIRBUS, Toulouse, France - ​2​Safety Data-CFH, Toulouse, France - ​3​HONEYWELL, Brno, Czech Republic {​estelle.e.delpech;alexandre.arnold;dominique.pronto}@airbus.com​, ​{laignelet;pimm;raynal}@safety-data.com​, michal.trzos@honeywell.com Abstract This paper describes the creation of the AIRBUS-ATC corpus, which is a real-life, French-accented speech corpus of Air Traffic Control (ATC) communications (message exchanged between pilots and controllers) intended to build a robust ATC speech recognition engine. The corpus is currently composed of 59 hours of transcribed English audio, along with linguistic and meta-data annotations. It is intended to reach 100 hours by the end of the project. We describe ATC speech specificities, how the audio is collected, transcribed and what techniques were used to ensure transcription quality while limiting transcription costs. A detailed description of the corpus content (speaker gender, accent, role, type of control, speech turn duration) is given. Finally, preliminary results obtained with state-of-the-art speech recognition techniques support the idea that accent-specific corpora will play a pivotal role in building robust ATC speech recognition applications. Keywords: speech corpus, spoken language, controlled language, air traffic control phraseology, speech recognition, accented English. 1. Introduction The AIRBUS-ATC corpus was developed in order to build a speech recognition system able to process Air Traffic Control (ATC) communications - messages exchanged between pilots and controllers using an English-based controlled language known as the ICAO phraseology (ICAO, 2007) - and Automatic Terminal Information Service messages (ATIS, airport information broadcast about weather conditions, available runways, etc.). The goal, in the long term, is to help pilots by giving them reliable ATC information in a persistent and visual way. Pilot: ​Montana, F-CD, request cancel my IFR flight, proceeding VFR estimating Borton at 1701 Controller: F-CD, IFR flight cancelled at 35, contact Montanan Information 125.750 Figure 1: Example of ICAO phraseology. ATC communications being very different from everyday conversations (see table 1), voluminous datasets like the SWITCHBOARD (Godfrey et al. 1992) and FISHER (Cieri et al. 2004) corpora cannot be used to build such a system although they were key in achieving human parity for conversational speech (Hannun et al. 2014; Xiong et al. 2016; Saon et al. 2017). SWB/FISHER ATC speech intelligibility good (phone) bad (radio transmission) accents US English diverse and non-native lexicon & syntax oral syntax everyday topics limited to ICAO phraseology or related speech rate standard high (Cauldwell, 2007) other - code switching Table 1: SWB & FISHER corpora vs. ATC speech The AIRBUS-ATC corpus intends to address this issue by providing a real-life corpus of transcribed English ATC messages spoken by non-native speakers. 2. Existing ATC speech corpora Six ATC speech corpora were found in the literature. Three of them are unavailable: the nnMTAC corpus (Pigeon et al. 2007) ​– ​24 hours of real-life, non-native military ATC, the VOCALISE dataset (Graglia et al. 2005) and the corpus of Lopez et al. (2013) ​– ​respectively 150h and 22h of real-life French-accented civil ATC. Available resources are the ATCOSIM (Hofbauer et al. 2008), …

Content truncated to 500 words