Strategies for improving low resource speech to text translation relying on pre-trained ASR models