Automatic Speech Recognition (ASR) has made significant progress thanks to the advent of deep neural networks (DNNs). In the context of under-resourced languages, for which few resources are available, spectacular achievements has been reported. ASR systems are a step forward for language documentation, as the annotation cost is considerably reduced for field linguists (manually annotated an audio file can take a tremendous amount of time), and the language is preserved and perpetuated through documentation. Previous `standard' deep neural networks reached very good performances for phonemic transcription (such as with Kaldi and ESPnet approaches).However, these methods only rely on the phoneme-level. In this thesis, we explore recently published ASR approaches which have shown to be effective on low-resource languages to produce word-level audio-aligned transcriptions. The first approach, based on self-supervised learning, is a speech model that uses a Connectionist Temporal Classification (CTC). The second, entitled wav2vec-U, proposes a framework intended to build an ASR system in a fully unsupervised fashion. With few resources at our disposal, we try to assess the usability that can be made from dictionaries. We conducted experiments on two low-resource corpora, the Yongning Na and the Japhug from the Pangloss Collection. The experimental results from the first approach demonstrate powerful word-level transcriptions with competitive error rates. Preliminary results are reported on the second approach. By a coverage measure of dictionaries on the available transcriptions, we show that these resources are not yet usable in the conducted approaches.