نبذة مختصرة : International audience ; The correct automatic identification and segmentation of phonemes is crucial for a more in-depth exploration of prosodic parameters on a syllabic level. As such, automatic phonemic transcription from spontaneous speech recordings has numerous applications, such as teaching or health monitoring. Such transcriptions are usually evaluated either in terms of correct phoneme estimation or temporal segmentation, each task being addressed by a dedicated system. However, no system to our knowledge has ever been evaluated on doing correctly the two tasks at the same time. This article evaluates a state-of-the-art Kaldi-based phonetic transcription system for spontaneous French. We use the Rhapsodie database, composed of spontaneous speech recordings with diverse levels of planning. Our phoneme recognition system obtains good results on phoneme and phoneme category identification (respective error rates of 19.2% and 13.4%), performed poorly on phonemes and category segmentation: an average of 40% of phoneme duration and 34% of phonetic categories duration have not been detected by it. On both metrics, the performances of the system increase with the degree of planning of the spontaneous speech. These results suggest that improvements are necessary for designing truly reliable automatic phonetic transcription systems to be useful for further analysis
Processing Request
No Comments.