The construction of text-to-speech (TTS) synthesis systems dates as far back as the industrial revolution, and is one of the most important natural language processing (NLP) tasks on the human-computer interaction spectrum. Numerous speech synthesis systems emerged within the past two decades; however, little has been done to extensively research and build such systems for low-resourced languages like Tigrinya. This paper presents a new large-scale speech corpus for Tigrinya, with 17.38 h of transcribed speech (11,639 utterances) of a single speaker. The dataset was trained on the Tacotron model for 600K steps, in which regardless of the training steps’ cap, generated fairly intelligible speech but with degraded natural prosody. The predicted speech scored an overall mean opinion score (MOS) of 3.41 when evaluated by experienced sound and audio signal processing experts, which is remarkable considering the unique intonational challenges present in Tigrinya.