Talk-to-text A.I. systems (TTS) are built to synthesize human speech so when a user interacts with it, the system can reply in a natural way that makes sense to the user. The job of a conversation designer (CXD) is to anticipate ways in which an A.I. TTS can accomplish this. Each language contains its own variety of dialects, and within those dialects each contains its own linguistic anomalies that present their own challenges for TTS and for the CXDs writing the strings.

Oftentimes languages contain pauses, intonations, or emphases that have an impact on the meaning of what’s being said. The English language, for example, has words that are spelled the same but sound differently when spoken, or sound the same when spoken but spelled different, and all carry different meaning depending on the context of what’s being said. These are called homonyms and they are split into two main categories: homographs and homophones. A homograph refers to words that are written the same but have different pronunciations (like bass- the fish, and bass- the guitar). A homophone sounds the same but is spelled differently (there, they’re, their).

Spoken language also sometimes requires dialectal adjustments to increase the naturalness of how the system responds to the user. For example, when American English-speakers say the words “let me” they often soften the articulation of those two words separately into one (lemme). This is called a soft articulate.

Currently there are no ways for a TTS system to know when to apply these linguistic anomalies.

I was tasked with creating a GUI tool to give CXDs the ability to fine-tune how the TTS output sounds to optimize naturalness, context or style in TTS performances. CXDs can insert pauses, add intonation/emphasis, change pronunciations or invoke different speakers or speaking styles.