Text to speech

The part of Amharic synthesis that is actually hard is not the voice, it is the text in front of it: a model that has only ever read clean prose cannot say a phone number, a price in birr or a time on the Ethiopian clock. Ours handles all three, because the normalisation rules were measured against real transcripts rather than inferred from a grammar.

What it does

• 01

  Amharic, Ge'ez numerals and Ethiopian clock times read as a speaker would

• 02

  Streaming output, so playback starts before synthesis finishes

• 03

  WAV and MP3, 24 kHz

• 04

  Deterministic for a given input, so a cached clip stays valid